WorksComputer Science
Self-Building Labs
Abstract
Self-driving laboratories (SDLs) accelerate scientific discovery by automating experimental execution and optimization. AI agents can increasingly propose new experiments, but existing SDLs can only run the procedures they were built for. Adapting them to a new experiment still requires extensive manual engineering, which makes laboratory flexibility the bottleneck for scaling discovery. We introduce the self-building laboratory (SBL), which designs, configures, and operates its own laboratory setup given a high-level scientific goal. SBL starts from whatever already exists, from an empty room to a working SDL, and builds the missing capabilities: the lab layout, its digital twin, robot skills, and experimental workflows. We demonstrate the system working across different robots and setups, scaling from a single workstation to laboratories with several robots working together. We validate SBL's end-to-end capabilities through a real-world quantum-dot synthesis experiment. By shifting human researchers from manually configuring and programming device automation to specifying scientific objectives, SBL makes scientific automation more accessible, flexible, and scalable. Project website: https://self-building-labs.github.io/
Website: self-building-labs.github.io
Date: November 16, 2026
Correspondence: [email protected], [email protected]
Introduction
Self-driving laboratories (SDLs) accelerate scientific research by automating experiments through the design–make–test–analyse (DMTA) cycle, in which experimental results inform subsequent decisions (1; 2; 3; 4; 5; 6; 7; 8; 9; 10; 11; 12. These approaches have been used to design biological systems, synthesize solid-state materials, optimize nanocrystals (13; 14; 15, to name a few. However, the autonomy of an SDL is typically bounded by the equipment and operational capabilities established during its development [16].
This boundary becomes a bottleneck as scientific research evolves. New research questions may require experimental procedures or instrumentation that were not anticipated when the platform was configured. Extending these capabilities typically requires specialists to integrate hardware and develop the software needed to execute new procedures [17, 12, 18]. Such modifications are time-consuming [19] and often depend on tacit knowledge that is difficult to preserve or transfer [20]. Changes in equipment placement or laboratory layout can further compromise the reliability of existing workflows. Modular platforms and robots that reconfigure equipment can adapt workflows using existing hardware and software [21, 4, 22]. Capabilities beyond those already available still require human development.
Recent advances provide an opportunity to address this limitation. Large language model (LLM)-based agents have demonstrated the ability to plan experiments, generate instrument control code, and interact with laboratory hardware [23, 24, 25, 26, 27, 28, 29, 30], while physics-based simulation enables developing digital twins of laboratory environments [31, 32]. Advances in robotics, particularly mobile manipulation, allow experiments to draw on instruments distributed across a laboratory by transferring samples between instruments [6, 33, 34, 35, 36, 37]. Language models can select existing robot skills based on their likelihood of success [38] and generate control programs that use sensory feedback [39]. Using robots as tools for general-purpose agents has also been discussed as a broader direction for robotics [40]. Recent studies test GPT-6-Astra for robot control in simulation [41, 42, 43]. Experiments and technical reports also show control of real robots, with some systems reusing experience or using a digital twin to guide actions [44, 45, 46]. These advances support flexible experimental execution, but adapting a laboratory to new scientific requirements still requires substantial time and expertise [47].
Here, we demonstrate a self-building laboratory (SBL) that incorporates the identification, development, and validation of task-required operational capabilities into the autonomous scientific process (Figure 1). Researchers define the scientific objectives that guide this process. SBL self-identifies gaps between these objectives and its available capabilities, then develops or extends the interfaces, skills, and workflows needed to address them. The system uses a digital twin to develop candidate operations and check them against task constraints before refining and validating them through physical interaction. Successful procedures are retained as reusable skills for subsequent tasks. Table 1 compares how experimental capabilities are developed and used across laboratory paradigms.
Figure 1. Concept of a self-building laboratory (SBL). Scientific goals guide the identification of capability gaps, development in a digital twin, and validation in the physical laboratory. Feedback and reuse support continued capability development across workflows, equipment, and entire laboratories. The image combines SDL2 at the Acceleration Consortium, University of Toronto, Canada, with its digital twin built by the SBL and rendered in Isaac Sim.
| Aspect | Traditional lab | Self-driving lab | Self-building lab (this work) |
| Experiment execution | Researcher-led procedures | Automated workflows using integrated capabilities | Workflows using capabilities developed as needed |
| New capabilities | Developed by researchers | Typically engineered by specialists | Developed and validated by agents |
| Knowledge retention | Written procedures and researcher experience | Protocols and script libraries tailored for one setup | Reusable skills across laboratories |
Table 1. Comparison of laboratory paradigms.
We examine SBLs capability development through digital twin construction, operational skill development, experimental workflow building, and transport prediction. In a quantum-dot synthesis study, SBL constructs a task-relevant digital twin and develops interfaces and feedback-guided skills that connect liquid handling, robotic plate transfer and optical measurement. On the existing RoboChem-Flex platform [11, 22], an agent establishes operating procedures for flow photochemistry and inline analysis under expert supervision. Separate transport studies on that platform develop and test a calibrated model for predicting downstream liquid and gas arrivals under different pump programmes. In simulation, the system also coordinates different robotic configurations and interconnected workstations. By incorporating capability development into the autonomous scientific process, SBLs could reduce the engineering effort required to adapt laboratory automation to new scientific objectives and make it more accessible to researchers without specialized automation expertise.
Results
Overview
SBL takes a user’s scientific goal together with a description of the current laboratory: photographs of the space, device identifiers, and any existing automation interfaces (Figure 2). The starting point can range from an empty room to a working SDL; in each case, SBL builds the capabilities missing for the goal: the lab layout, its digital twin, robot skills, and experimental workflows (Table 2). It builds a digital twin of the laboratory, from individual device models to the layout of the whole space, using robot-mounted cameras where available to gather additional observations (§ 2.2). SBL identifies the capabilities the desired workflow requires and acquires the missing skills using the digital twin and, with user approval, trial-and-error in the real world (§ 2.3). This allows it to go from a user’s scientific goal to safely executing the desired workflow in the real world, with minimal human intervention (§ 2.4). The RoboChem-Flex study examines how an agent develops operating procedures and predictive transport capabilities on an existing automated platform (§ 2.6). The same process extends to designing laboratory layouts and coordinating several robots across workstations (§ 2.8).
| Starting point | Already exists | SBL builds |
| Empty laboratory | N/A | Layout, digital twin, skills, and workflows |
| Existing manual laboratory | Equipment | Digital twin, skills, and workflows |
| Automated laboratory requiring reconfiguration | Equipment and some skills | New layout, updated digital twin, and adapted skills |
| SDL with a new protocol | Equipment and skills | New workflows and any missing skills |
Table 2. Starting points for a self-building laboratory. The same process applies in each case; only the required set of capabilities differs.
Figure 2. Architecture of the self-building laboratory. The scientific agent translates user goals into experimental plans, while the twin agent constructs a digital twin, develops task-relevant skills, and validates workflows in simulation. The safety agent uses the twin for scenario analysis and generates runtime guardrails through which all physical laboratory commands pass. The scientific agent invokes the physical laboratory through this layer. Physical state feedback supports twin calibration and subsequent simulation, while simulation and physical execution traces support skill distillation and retain experience as agent memory. Reusable skills and memories inform subsequent tasks.
Building digital twin
Figure 3 shows how the twin agent reconstructs individual equipment and assembles these models into a calibrated laboratory scene. The user provides laboratory photographs, equipment model numbers when available, and device application programming interface (API) endpoints. From these inputs, the agent builds a task-oriented representation of equipment identity, geometry, articulation, spatial arrangement, and control interfaces.

Figure 3. Building a digital twin of the laboratory. (a) Starting from laboratory images and equipment model numbers when available, the twin agent identifies individual devices, searches product documentation, and retrieves or constructs equipment models with task-relevant geometry, articulation, and control interfaces. (b) These models are assembled into an initial simulation scene and refined using colour and depth observations acquired from multiple robot poses. The agent compares observed and rendered views and iteratively updates equipment poses to align the scene with the physical laboratory.
Equipment reconstruction begins with identifying the individual devices from the photographs and any supplied model numbers. For each device, the agent first searches for existing computer-aided design (CAD) models and Unified Robot Description Format (URDF) files. If these are unavailable, the agent reconstructs an approximate 3D model using manufacturer documentation, technical specifications, and reference images. The agent first estimates the device dimensions using available measurements and known reference objects, then breaks down the device into individual components, such as frames, housings, doors, and moving mechanisms. Using Python, the agent constructs these components from basic geometric primitives and custom meshes. The agent then iteratively refines the generated models by comparing rendered views with reference images and adjusting geometry, proportions, and relative positions. Once the geometry is established, the agent adds articulations for movable components by defining their joints, motion axes, and limits. This process produces simulation-ready equipment models without requiring manual CAD modelling. Missing or uncertain parameters are recorded for further refinement.
The agent then assembles the equipment models into an initial scene based on the observed laboratory layout. To refine this arrangement, the agent plans camera poses and acquires colour and depth observations from
each pose using its wrist-mounted camera. The agent compares observed and rendered views and iteratively updates equipment positions and orientations. This procedure reduces equipment-pose discrepancies and establishes the robot-to-device poses needed for motion planning.
To support operations within the reconstructed scene, the agent implements simulated devices with the same API interfaces as their physical counterparts but different endpoints. Selecting the target endpoint determines whether an operation runs in the twin or the physical laboratory. A monitoring system polls physical device status through the APIs and updates the corresponding states in the twin. Using the calibrated scene, the agent plans robot motion in the twin and sends the resulting trajectories to the physical execution endpoint to control the robot. The twin remains a living replica of the physical laboratory, continually updated as new observations become available.
Building laboratory skills
Figure 4 illustrates laboratory skills at three levels: device execution interfaces, embodied operations, and scientific procedures and interpretation. Scientific requirements create demand for embodied skills, which in turn depend on the capabilities exposed by device interfaces. As these requirements arise, the agent develops missing skills and interfaces, building the operational capabilities needed to complete the task.
Figure 4. Scientific tasks drive laboratory skill development. A plate-sealing task drives the development and integration of embodied and device skills. Digital-twin planning and feedback from physical execution guide refinement of the twin and skills, connecting scientific requirements to executable laboratory capabilities.
Physical skills that enable the robot to perform experimental tasks are developed by the agent in simulation and on the real robot. In simulation, the agent maps the required experimental operations onto available equipment and explores candidate action sequences through trial and error. Specifically, using object states, the agent automatically defines grasping and placement poses, then uses a motion planner [48] to generate collision-free trajectories through the cluttered laboratory workspace. It composes these motion segments into an experimental workflow.
Directly transferring a workflow from simulation to the real world is limited by calibration and modelling errors. Transport and coarse localization can often be directly deployed. In contrast, precise interactions, like grasping and insertion that require millimetre-scale precision, are more sensitive to these errors. There, online exploration is needed. The agent can autonomously use local sensory feedback to refine those interactions on the real robot. During exploration, it selects actions based on visual observations, depth measurements, robot state, and force/torque feedback, then examines the resulting sensory responses to identify successful strategies.
Through this process, it learns to interpret feedback and use it to guide subsequent actions. Placement into a constrained fixture, for example, can involve visual alignment, guarded descent, contact detection, lateral adjustment, and resumed insertion after the measured load decreases.
A particularly challenging example is learning to operate an high-performance liquid chromatography (HPLC) instrument originally designed for manual operation. Loading a well plate into its holder requires sub-millimetre positioning accuracy, while opening and closing the drawer involve sustained contact and constrained mechanical motion. The robot must engage a spring-loaded locking mechanism and pull the drawer along its axis. Before planning subsequent actions, it must determine how far the drawer has opened and where the holder is now located. Limited camera visibility, narrow clearances, and the curved inner surface of the drawer further constrain the choice of contact points and manipulation trajectories. The robot must also disengage without applying excessive force or torque. To address these challenges, the agent modelled the drawer’s articulation in the digital twin, explored alternative manipulation strategies, and refined the model and action sequence based on observations from physical trials.
The agent can complete a range of tasks using this approach through repeated cycles of reasoning, code generation, physical action, and observation. However, these cycles were too slow to meet the requirements of a chemical workflow. To reuse the experience gained during exploration, the system distills successful action sequences and their associated feedback into executable skills. Once an agent establishes a repeatable procedure, it encodes it as a Python script containing deterministic actions, sensor-based conditions, and recovery branches, then tests it through repeated physical execution. The resulting skill can then run without agent involvement during nominal operation. Agentic reasoning resumes only when execution departs from the conditions encoded with the skill.
Safety Guardrail. Safety is a critical consideration when deploying autonomous agents in physical laboratories, particularly in environments shared with human researchers [49, 50, 51]. To address this challenge, SBL incorporates safety mechanisms at both the simulation and physical execution levels. Specifically, beyond collision-free motion planning, SBL enables safety discovery by leveraging artificial intelligence (AI) agents to autonomously propose and evaluate potentially hazardous scenarios in simulation. Through large-scale exploration of possible failure conditions, we demonstrate that SBL identifies potential risks and generates candidate safety guidelines that can be incorporated into subsequent workflow planning. These findings also provide a foundation for building a shared knowledge base of laboratory hazards and safety practices. During physical execution, SBL implements multiple layers of safety guardrails. A separate control interface allows human operators to stop the entire experimental workflow at any time. Additionally, SBL enforces speed limits on robot trajectories and continuously monitors onboard force sensors to detect unexpected contact and trigger protective stops. Together, these capabilities extend SBL beyond conventional safety enforcement toward proactive hazard discovery and risk-aware autonomous laboratory operation.
Building experimental workflow
Figure 5 demonstrates the quantum dot synthesis experiment, which combines Opentrons liquid handling, plate sealing, and photophysics measurements. The experiment highlights the challenges of end-to-end task planning and long-horizon robotic execution. A rail-mounted robotic arm transfers a 96-well plate between instruments, requiring millimetre-level precision during pick-up and insertion while keeping the plate level throughout each trajectory. SBL autonomously performs these tasks after a human operator provides the initial information and approves the proposed plan.
Figure 5. SBL-coordinated CdS synthesis and optical readout. SBL executed 236 OT-2 commands to prepare a 96-well plate, followed by 19 robotic stages for transfer, sealing, re-grasping, and reader loading. Overlaid video frames and teal arrows illustrate plate handling. The heatmap shows raw fluorescence intensity at 560 nm after 8 h, with well H8 highlighted. The adjacent plots show the evolution of the H8 emission spectrum and its fluorescence intensity at 560 nm alongside the temperature trace.
SBL used read-only API calls to establish the deck layout, pipettes, tip inventory, and sealer and reader capabilities, then adapted the requested chemistry to these constraints. It wrote and verified a standalone Opentrons Python protocol offline, ran it in an Opentrons simulator matching the robot’s software version, and froze the reviewed file by recording its SHA-256 hash. The protocol consists of 236 commands covering setup, flow rates, tip, and liquid commands, all with timestamped event markers. These actions were validated against the gateway’s live OpenAPI schema, with the gateway configured to reject any change to the protocol’s hash. Before execution, SBL re-read the protocol and required the equipment to be ready and idle. It then sent the actions sequentially, recording device state before each request and confirming afterwards that no motion was in progress. For the subsequent plate transfers, SBL used the digital twin for motion planning,
and visual and sensory feedback to guide grasping and placement. The robotic arm first transferred the plate from the OT-2 to the PlateLoc for sealing. To prepare the plate for insertion into the Cytation 5, SBL then used an intermediate plate holder as a transfer station, where the arm re-grasped the plate along its short side. The arm placed the plate in the reader’s plate slot, and SBL registered it with the Cytation 5 for optical readout.
The experiment synthesized cadmium sulfide quantum dots ( QDs) in a 96-well microplate, using and thiourea as precursors and 3-mercaptopropionic acid (MPA) as the stabilizing ligand [52, 53]. At a fixed Cd concentration, the screen varied the MPA:Cd molar ratio (0, 1, 1.5 and 2), the thiourea:Cd molar ratio (0.75, 1.5 and 3), and the buffer-stock pH (8.5, 9.0, 9.5 and 10.0). The plate was incubated at 60 °C in a microplate reader, which monitored absorbance and photoluminescence over 8 h.
After 8 h, the strongest fluorescence at 560 nm was observed in the pH 10 rows at MPA:Cd ratios of 1 and 1.5, with the highest intensity in well H8. In H8, the emission band above 500 nm intensified through the final measurement, consistent with progressive growth of MPA-capped QDs [52]. The complete plate-reader spectra are provided as an HTML file in Data S1.
Building capabilities on an existing platform
SBL also combined the previously developed HPLC-manipulation skills with plate transport to load a deep well plate from the OT-2 into an HPLC for sample analysis. Figure 6 illustrates the resulting workflow, with successive motions planned around drawer-position changes and whether the plate was held or released. The robot carried out the sequence, integrating free-space transport, sustained contact with the drawer, and plate release within the holder’s narrow clearance. This example showed how instrument-specific skills could be composed into a workflow connecting liquid handling with sample loading for analysis.
Figure 6. Robot-assisted plate loading into an HPLC. Teal arrows indicate the plate-transfer direction. Selected video frames illustrate five workflow phases. The 32-action workflow comprises 28 arm motions and 4 gripper actions.
Building capabilities on an existing platform
Figure 7 examined how SBL could develop operational and predictive capabilities on an existing SDL. RoboChem-Flex combines liquid handling, irradiation and inline analysis in a modular flow-photochemistry platform [11, 22]. Starting from the assembled apparatus, its publication and supporting files, low-level device APIs, photographs and videos, SBL built a digital twin and developed a model of the 28 published fluidic channels in approximately 90 min (7a). The fluidic model includes tracked routes, syringe inventories and the sequence of liquid and gas segments. This model was iteratively developed with physical trials to predict fluid transport, thereby establishing reusable procedures for priming and transfers (7b).
Figure 7. Building capabilities on RoboChem-Flex. (a) Digital-twin reconstruction from supplied resources, with four paired camera and rendered views. (b) Needle descent into vial F-C3 as iteratively tried across simulation and real image observations at controller depths; the final position was only simulated as it approached the assumed vial floor. (c) Digital twin internal flow path model with calibrated (blue numbers) and measured (orange numbers) inter-sensor distances and observed (orange) and predicted (blue, dashed) phase signals, with high and low signals denoting liquid and gas. Predictions propagate boundaries detected at H using inter-sensor displacements calibrated on five packets. For a 650 acetonitrile packet at 0.78 (left), outlet arrival errors are within 1.3 s. An agent-designed 45 s pause (right) gives an outlet liquid window of 94.5 s versus 94.6 s predicted. Both tests use the lamp off; the late pulse is chase liquid. (d) Reaction scheme, corrected product-to-aromatic integral ratios for seven eligible runs versus nominal residence time, and baseline-corrected NMR spectrum of annotated run. Purple and green bands mark aromatic and product integration windows; the product region is expanded tenfold.
The agent calibrated fluid transport from sensor measurements and used the model to calculate downstream arrival times for different pump programmes. The same calibration was used to predict programmes beyond those used to fit the model. Across twelve held-out transits lasting 5–10 min, the observed transit midpoints differed from the predictions by less than 1.4 s, on the same order of magnitude as the response delay from device API. These tests used acetonitrile in liquid-filled and nitrogen-cleared paths with the lamp off.
The model also predicted liquid and gas arrivals for programmes that accelerated flow, paused near the outlet or slowed flow during passage (7c). A programme with a 45 s pause produced an outlet liquid window of 94.48 s compared with 94.60 s predicted. These comparisons show how the transport model can evaluate pump programmes for coordinating reactor passage and downstream handling.
The agent combined the handling procedures into a ten-run photochemistry campaign. The reaction involved adding tetrahydrofuran (THF) to benzylidenemalononitrile using tetrabutylammonium decatungstate (TBADT) as a photocatalyst [54, 55]. It passed reaction packets through the reactor under 370 nm irradiation and positioned them for inline nuclear magnetic resonance (NMR) acquisition (7d). The campaign varied THF loading, catalyst loading and nominal residence time. The seven full-power measurements included in the comparison gave product-to-aromatic integral ratios of 0.30–0.36.
Building new laboratory hardware
Figure 8 demonstrates how SBL can build and install a missing hardware component to support a new experimental workflow. LumaStir was originally designed in-house to illuminate and stir samples in vials, the details of which will be published elsewhere. Stirring the sample after solid dosing required a holder to keep the vial in position once the robot placed it. SBL designed a three-position holder with openings sized for the sample vials and a base that fit the empty slot in the front row of LumaStir (Figure 8a). It sliced the holder geometry and printed the part on a Bambu Lab A1.
Figure 8. Building and using a vial holder on LumaStir. (a) The assembly model shows the empty front slot and the new holder (teal). (b) The wrist-camera image compares the predicted outline of the holder on the print bed (dashed) and the outline fitted to its top edges (solid) before UR5e pickup. (c) The wrist view guided a lateral adjustment above the slot, and the closed gripper pressed the holder into place after release. (d) The digital twin motion plan guides the handoff of a single vial from the UR3e (left in the photograph) to the UR5e (right) in real world. (e) The installed holder supports the vial in the rightmost opening. The liquid surface is visible through the front window before illumination and during stirring.
In the digital twin, SBL used the print layout and the bed’s programmed final position to predict the holder’s pose at the end of the print. After printing, SBL refined the holder’s pose estimate using a wrist-camera image and added a short motion to join the planned UR5e trajectories for transfer and placement in the front slot. The UR5e lifted the holder from the print bed and positioned it above the front slot (Figure 8b,Figure 8c). SBL checked the holder’s alignment with the slot using the wrist camera and made a small lateral adjustment before lowering it. The arm adjusted the holder’s angle during insertion, then released it and pressed it into place with its closed gripper.
SBL planned the bimanual vial handoff in the digital twin. At the bench, the UR3e passed a prepared vial to the UR5e (Figure 8d and Video S3). The UR5e placed the vial in the rightmost opening of the new holder and withdrew. The agent switched on the light at that position and ran the stirring motor for 60 s. Through the holder’s front window, we observed a vortex form in the illuminated vial (Figure 8e).
Building scalable labs
One major challenge in laboratory automation is scalability, as existing systems are often tailored to specific hardware configurations and experimental workflows ([17, 19]). SBL addresses this challenge by automatically designing, evaluating, and optimizing laboratory configurations in simulation before physical deployment. We demonstrate its scalability across three scenarios: layout and motion optimization in a confined space, coordination of heterogeneous robots across multiple workstations, and sample transport across rooms and floors of a laboratory building. Across these scales, SBL adapts laboratory layouts, robotic configurations, and experimental workflows to different spatial and operational requirements.
Scaling layout design through large-scale parallel simulation. To design an air-sensitive flow chemistry workflow that involves solid dispensing in a confined glovebox, SBL jointly explored equipment placement and robot motion. SBL first placed the existing RoboChem-Flex digital twin in the glovebox model, with its sampler and fraction collector on the worktop, and added a UR3e, a solid dispenser and a shaker. The workflow used one process vial for solid dispensing, liquid addition and mixing before injection into a flow photoreactor. SBL explored alternative layouts, including one with the sampler rotated by and the fraction collector elevated to use the glovebox’s vertical space. In this layout, SBL used parallel geometric screening to evaluate 10,000 distinct combinations of sampler position and vial grasp angle in ten rounds of 1,000 candidates (see Video S4). From these candidates, SBL selected a new position for the sampler in the glovebox model. It then refined robot transfer routes for both sampler positions and compared the workflow variants. Digital-twin search connects equipment placement, robotic access and motion time, enabling laboratory configurations to be explored and refined around a new workflow before deployment.
Scaling across multiple robots and workstations. We demonstrate the scalability of SBL through an end-to-end buffer preparation workflow involving a solid-handling workstation, a wheeled mobile manipulator, and a liquid-handling workstation, as shown in Figure 9. Given only the high-level experimental goal, SBL autonomously reconstructs digital twins of the workstations from physical images, including particle-based simulation of solid dispensing, and generates coordinated actions across heterogeneous robotic platforms. The workflow begins with dispensing solid materials into six vials at the UR workstation. Since the vial and rack designs are unspecified, SBL automatically generates CAD models that ensure secure vial placement and the ability for robot to grasp. Due to a central shelf obstructing direct rack transportation, SBL plans a collision-free handoff between two UR robots to transfer the rack across the workbench. SBL then coordinates a wheeled mobile manipulator to transport the rack to the xArm workstation, where the rack is placed into an Opentrons liquid handler for reagent dispensing and subsequently transferred to a shaker for mixing. This demonstration highlights SBL’s ability to scale laboratory automation across heterogeneous robots and distributed workstations by jointly coordinating robotic manipulation, inter-workstation transportation, and instrument operation within a unified workflow.
Figure 9. Scaling self-building laboratories across robots and workstations. The reconstructed simulation layout (top) integrates a UR3/UR5 solid-handling workstation, a mobile manipulator, and an xArm-based liquid-handling workstation. Given a high-level buffer-preparation goal, SBL generates a coordinated workflow (bottom) in simulation and task-relevant powder and fluid simulation models for solid and liquid handling.
Scaling to multi-floor buildings. We envision automated laboratories scaling to entire buildings, where mobile robots connect robotic workstations with shared instrumentation by transporting samples across rooms and floors. In typical chemistry buildings, instruments such as X-ray diffractometers and NMR spectrometers are housed in dedicated facilities shared by multiple laboratories, requiring chemists to manually transport samples for characterization. To demonstrate SBL’s scalability to this setting, we simulate an NMR sample preparation and characterization workflow in the new Acceleration Consortium building (see Video S5), which is currently under construction. Starting from a building schematic, we ask SBL to add lighting and laboratory equipment and to configure articulated doors and elevators. In the simulated workflow, an Opentrons liquid-handling robot prepares an NMR sample, after which the tube is shaken to move the liquid to the bottom and then capped. A mobile robot transports the sample through corridors and takes an elevator to the NMR facility, where the tube is placed in a holder and loaded into the spectrometer. This demonstration points toward a broader vision in which an entire laboratory building operates as a coordinated, autonomous discovery system. Before construction is complete, SBL enables researchers to design, test, and optimize collaboration among robots, laboratories, and shared instruments, laying the groundwork for building-wide orchestration of experiments and advancing materials discovery from individual automated laboratories to fully autonomous laboratory buildings.
Discussion
SBL brings experimental capabilities into the scientific process. In the quantum-dot study, this meant adding missing controls, such as a joint-servoing interface for the arm, and establishing the conditions under which an operation succeeds, such as feedback-guided insertion when scene calibration was not precise enough. The RoboChem-Flex study extends this process to operating procedures and a calibrated model for predicting fluid transport on an existing automated platform. Both studies developed capabilities through physical interaction under human supervision. The 3D printing case illustrates how SBL can use experimental requirements to identify a missing hardware component and fabricate it. We examined six-vial rack transfers and transport between workstations in simulation.
What the agent can build depends on what the laboratory infrastructure lets it observe and verify [56]. Some device failures produced no error from the API, leaving the agent unable to tell success from failure [57]. Fine insertion became possible only after force and torque sensing exposed contact conditions that cameras and scene calibration could not [58]. On RoboChem-Flex, optical phase measurements made it possible to check transport predictions against observed fluid motion.
RoboChem-Flex shows how knowledge from physical trials can inform subsequent experimental planning. A calibrated fluidic transport model allows candidate pump programmes to be evaluated before execution, with subsequent measurements testing the predictions and informing model refinement. The separate exploratory flow model could support comparisons of equipment configurations by explicitly representing tube geometry and gas dynamics, subject to calibration and physical validation.
The digital twin also supports safety. Simulation supports trajectory planning, collision checks, and exploration of hazardous scenarios before physical execution. The resulting constraints inform workflow planning and runtime guardrails, including robot speed limits and force-triggered protective stops. These guardrails must not rely on the model refusing unsafe instructions: tests on real robots show that GPT-6-Astra and other models will follow them [59]. The validity of twin-based checks depends on the geometry, device states, and physical interactions represented in the twin [60]. Collision checking covers modelled geometry; chemical hazards and uncertain contact conditions, such as the millimetre-scale grasping and placement in our workflow, require additional checks during physical execution. Safe capability development therefore requires both space for exploration and enforced limits on physical actions, with execution observations used to refine the twin and its constraints [49, 51].
Human involvement remains necessary as new operating conditions emerge during long-horizon tasks [24, 12]. Agents retrieve documentation, reconcile information, and execute code at a pace that makes continuous human inspection impractical. During development, we observed both actions based on assumptions that required clarification and overly cautious motion decisions that obstructed task progress. Disagreements between agents also produced repeated exchanges that required human intervention to resolve. These experiences identify a communication requirement: the system must surface the assumptions, conflicting evidence, and operational consequences relevant to a decision. Researchers retain responsibility for scientific objectives and supply physical facts that the available interfaces cannot observe. On RoboChem-Flex, operator guidance helped establish a needle working depth that camera observations alone could not determine. Twin-based checks and runtime guardrails provide a basis for delegating routine verification, while unresolved conditions bring the researcher enough context for intervention. This division directs human expertise towards the unresolved decisions that limit further capability development. This need for expert involvement, in turn, calls for training researchers to work in SBL environments, where human–agent–robot interaction becomes central, in contrast to both traditional human-led discovery and established SDLs running predefined workflows.
For this development process to accumulate capability, successful exploration must produce procedures that can be reused [61, 62]. Because agentic cycles were too slow for the chemical workflow, successful explorations were distilled into executable policies that run without the language model and return control to the agent when conditions fall outside their scope. Encoding these conditions makes operational knowledge reusable beyond the original exploration. Reuse across changed equipment or layouts requires checking those conditions and validating the adapted procedure [63]. Related work tests reusable skills with visual feedback in simulation and on real robots [44], and programs that check their own outcomes in simulation [64]. Capability development thus becomes part of the scientific process rather than a prerequisite. The next step is to measure whether successive cycles improve task performance and reduce the engineering effort and human intervention needed for new tasks, while retaining validated capabilities. This would provide a quantitative basis for assessing self-improvement in physical AI.
A global community of researchers can extend this process across laboratories by sharing the resources needed to develop and assess new capabilities [65, 66, 12, 67, 68]. Shareable resources include equipment models and digital-twin assets, device interfaces, operational skills, execution records, and safety guidelines, along with their applicability and validation conditions. Shared safety standards would provide a common basis for assessing these resources, while deployment in another laboratory would still require local checks [51]. A global capability map of SDLs would connect scientific objectives to available instruments, skills, and resources, and expose the gaps that further capability development must address [12, 65]. Such a map could support planning investigations that draw on capabilities distributed across laboratories. The result is a dynamic ecosystem for SBLs, in which scientific infrastructure evolves alongside scientific inquiry. The case for self-building laboratories is strongest where engineers cannot follow the equipment: remote field stations, contaminated or hazardous sites, and even laboratories in space, which must operate under constraints on resources, communication, and crew time and develop new capabilities from what is already on board [69, 70].
Methods
Software We used openai-gpt6-astra to construct and modify laboratory representations, generate experimental workflows, and invoke planning and equipment-control tools. The digital twin was implemented in NVIDIA Isaac Sim 6.1 using Universal Scene Description (USD) [71]. Equipment representations were assembled from available CAD models, reconstructed geometry, laboratory photographs, and device documentation. Articulated components were represented using joints, while liquid-handling states tracked liquid volume and composition. Robot motion planning used cuRobo [48] with robot collision-sphere representations and environment obstacles represented by meshes and cuboids. Collision checking was performed within the GPU planning pipeline. During workflow execution in the physical laboratory, all commands were issued through equipment-specific APIs, and robot and instrument status were recorded during operation.
Acceleration Consortium SDL2 equipment The experimental platform comprised a five-degree-of-freedom xArm5 robotic manipulator mounted on a motorized linear rail and equipped with a parallel gripper (BIO GRIPPER G2) and a 6-axis force/torque sensor for microplate handling (UFACTORY). The rail extended the robot’s access across the liquid-handling, sealing, and measurement stations. Liquid dispensing was performed using an Opentrons OT-2 pipetting robot equipped with an eight-channel pipette (P300 Gen2) on the left, and a single-channel pipette (P1000 Gen2) on the right. The microplates were sealed using an Agilent PlateLoc Thermal Microplate Sealer and characterized using a BioTek Cytation5 multimode reader. We used 19-mm-tall 96-well flat-bottom (Agilent 201242-100) and 45-mm-tall deep 96-well U-shaped bottom (Agilent 5043-9302) polypropylene microplates. Intel RealSense D435i and D405 RGB-D cameras provided laboratory observations and close-range views of manipulation interfaces. Each instrument was exposed as an authenticated HTTP service with a read-only GET /status endpoint, an OpenAPI schema, and documentation of available functions and valid input ranges. Measurement responses were saved incrementally as CSV files, together with run metadata. All machines connect through a private Tailscale network, with REST endpoints provided for instrument access. The central API checks the agent’s authorization and records each device claim.
Bimanual transfer and 3D printing The bimanual setup included a six-degree-of-freedom UR3e robotic arm (Universal Robots) equipped with a two-finger gripper (Robotiq 2F-85) and a wrist-mounted Intel RealSense D435i RGB-D camera. A six-degree-of-freedom UR5e robotic arm (Universal Robots) was equipped with a two-finger gripper (Robotiq 2F-140) and a wrist-mounted D435i RGB-D camera. The in-house LumaStir device accommodated two rows of three vials for illumination and stirring. The vial holder was fabricated from poly(lactic acid) using a Bambu Lab A1 3D printer fitted with a 0.4 mm nozzle.
Quantum-dot synthesis Chemicals were purchased from Sigma-Aldrich and used as-is without further purification. A series of pH 8.5, 9, 9.5 and 10 buffer solutions at 1 M concentration were prepared from 1 M and 1 M stock solutions with volumetric ratios of 17.0:3.0, 12.8:7.2, 7.2:12.8, 3.0:17.0 (mL:mL), respectively. The 0.01 M , 3-mercaptopropionic acid (MPA) and thiourea solutions were prepared directly in deionized . All operations were performed in fume hoods, and appropriate PPE was used for manual stock preparation. The solutions were transferred to the corresponding wells of a 45 mm 96-well deep-well source plate using volumetric pipettes. The source plate was then placed on the OT-2 pipetting robot for downstream reactions. Column 1 held buffers (1.5 mL): rows A–B pH 8.5, C–D 9.0, E–F 9.5, G–H 10.0. Column 2 held 10 mM , column 3 10 mM MPA and column 4 10 mM thiourea. Column 5 held deionized water (1.5 mL). The absorption values were measured at 350, 400, and 450 nm wavelengths. The emission spectra were measured in the 450–680 nm range and 10 nm intervals using 360 nm excitation wavelength.
RoboChem-Flex equipment The experimental platform [22] comprised a three-axis sampler with main and auxiliary needles, a three-axis collector, 10 mL and 1 mL syringe pumps, an injection port and sample loop, and a mixing chamber. The flow path included an approximately 3.9 mL perfluoroalkoxy-tubing photoreactor in a 370 nm lamp housing, a back-pressure regulator, a nitrogen line and a waste/analysis valve. Inline analysis used a benchtop NMR spectrometer (Spinsolve, 61.8 MHz). Five optical phase sensors monitored fluid transport, and four cameras were added for this study. Equipment control used version 3 of the device API, which exposed device primitives and observations but did not supply the deck map. Specified operating limits were 1 mL for the pumps and 1000 mm for the gantries.
Flow photochemistry Stocks were benzylidenemalononitrile at 2.0 M, TBADT at 17 mM and neat THF, with acetonitrile as diluent and carrier. Each 650 L packet contained substrate at a nominal 75 mM. Stocks were combined in a mixing vial with seven reciprocating 175 L sampler strokes and injected as a gas-capped packet. Reaction conditions varied THF loading, catalyst loading and nominal residence time. Pump rates of 0.26, 0.39 and 0.78 mL corresponded to nominal reactor residence times of 15, 10 and 5 min. The eligible comparison set used full lamp power. Each packet was positioned in the NMR flow cell and measured using WET solvent suppression [72] on acetonitrile, with 64 scans, 3.3 s acquisition time and 20 s repetition time. The reaction readout was the ratio of the product-window integral (3.0–4.0 ppm) to the aromatic-window integral (5.7–7.1 ppm) on the instrument’s unreferenced scale. Each integral used a linear baseline fitted to flanking windows.
Acknowledgments
J.B. acknowledges funding from the Eric and Wendy Schmidt AI in Science Postdoctoral Fellowship Program, a program by Schmidt Futures. E.S. and T.N. thank the European Union Innovation Council for the funding of ReaCtor (101099405). A.A.-G. thanks Anders G. Frøseth for his generous support. A.A.-G. also acknowledges the generous support of Natural Resources Canada and the Canada 150 Research Chairs program. This work was supported by the AI2050 program of Schmidt Sciences. This research is part of the University of Toronto’s Acceleration Consortium, which receives funding from the CFREF-2022-00042 Canada First Research Excellence Fund. This work was supported by the Defense Advanced Research Projects Agency (DARPA) under Agreement No. HR0011262E022.
Digital twin
Custom plate holders
Custom holders provided plate interfaces at the Opentrons, shaker, optical-table deck, PlateLoc and HPLC drawer (Figure S1). The optical-table deck holder was also used near the shaker and rail. The CAD models show the holder geometry used to represent these interfaces in the digital twin, as well as the 3D-printed and installed holders in the real lab.
Figure S1. Custom plate-holder CAD models. Holders for the a Opentrons, b shaker, c optical-table deck, d PlateLoc and e HPLC drawer. Models are shown at individual scales to make their features visible.
Motion planning
The agent planned robot transfers in the digital twin using the observed robot state, equipment geometry and task-specific start and target poses. The planning scene included the gripper and, during transport, the held plate and its pose relative to the tool. The agent used cuRobo [48] to generate joint trajectories and checked the robot, tool and plate against the surrounding geometry. For plate transfers, it also checked the plate orientation along the planned trajectory.
The xArm and rail moved in separate segments: the rail remained fixed during arm motion, and the arm remained fixed during rail motion. The agent checked trajectory limits and controller interpolation before execution, then transferred control at stopped positions between segments. At instrument interfaces, local grasping and placement used visual and force/torque feedback to refine the interaction.
Joint-trajectory execution limits
The servoJ interface enforced joint-space limits on every streamed trajectory and two stop behaviours (Table S1). A constrained stop decelerated the arm along its path; a hard stop triggered the controller emergency stop. Collision checking was performed during planning in the digital twin; the trajectory executor itself checked only joint-space limits. The controller streamed joint state at about 100 Hz, and each session log recorded every command sent and every pose reported.
| Parameter | Value |
| Controller mode | 1 (servo); position moves use mode 0 |
| Command rate | 100 Hz default; 50–200 Hz per session |
| Interpolation | Piecewise Hermite: quintic (C2) with velocities and accelerations, cubic (C1) otherwise |
| Lead-in | Rest-to-rest minimum-jerk quintic at 50 % of speed and acceleration limits |
| Joint acceleration limit | 500 ° s |
| Joint speed limit | 180 ° s × safety-level factor (0.5 at the default level, i.e. 90 ° s) |
| Start / settle tolerance | 0.5° / 0.2° within 3.0 s |
| Trajectory size | ≤600 s, ≤60 000 points, ≤5 000 points per uploaded chunk |
| Per-tick check | Step ≤ 1.2 ; implied acceleration bounded |
| Timing | Absolute deadlines, 5 ms sleep slices, garbage collection paused; re-anchor when >0.5 period late; stop when >5 periods late |
| Constrained stop | Time-scaled deceleration along the path, ≤2.0 s; triggered by cancel, claim loss or timing fault |
| Hard stop | Emergency stop on any SDK error, controller error or warning, unexpected state, or leaving mode 1 |
| Claim watchdog | Claim and mode re-checked every 0.1 s; claim TTL 30 s, heartbeat 10 s |
| Bench accuracy at 100 Hz | Lag 15 ms; tracking error ≤0.11°; final error <0.001° (mode 6: 65–95 ms, 0.73–0.76°) |
| Scope | Joint space only; no collision or Cartesian checking in the trajectory path |
Table S1. Joint-trajectory (servoJ) execution parameters on the xArm5.
Feedback-guided plate placement
The agent estimated the placement target from holder geometry, wrist-camera observations and the robot and gripper state (Figure S2). It aligned the held plate with the holder and corrected its orientation in free space when required, accounting for the displacement between the tool and plate.
Figure S2. Information used in feedback-guided plate placement. a Saved Isaac Sim view of the PlateLoc scene, shown for context. b The procedure combined holder geometry, camera observations, robot and gripper state, and force/torque feedback to guide alignment, contact search and release. The scene shown here is for illustrative purposes only.
The placement script monitored force and torque during descent. When contact occurred above the expected seating height, it held the plate and performed bounded lateral probes to find a position that reduced the contact load before continuing descent. The script evaluated height and contact-response conditions before opening the gripper and withdrawing vertically. After release, camera observations were used to check that the plate remained in the holder and that the gripper had cleared it.
Safety guardrails
Two simulated scenarios evaluated checks on plate transport and local manipulation (Figure S3). Before a transfer, the workflow checked the gripper opening and object-detection state against the expected grasp ( mm).

Figure S3. Simulated safety guardrails. a An illustrated misaligned grasp at the PlateLoc, with the plate raised on one side; the gripper-state check holds a transfer when the opening or object-detection state is inconsistent with the expected grasp. b A finger contacting the HPLC drawer during pulling out; force monitoring requests a stop when an unexpected load exceeds the threshold. The plate and drawer poses illustrate the failure scenarios in the digital twin; the arrow shows the downward motion and the marker indicates loading at the drawer support. The force simulation used a motion speed of , a contact stiffness of , a damping coefficient of , a monitoring interval and a stop latency.
During local manipulation, the workflow monitored wrist force and torque for unexpected loading. In a simulated obstruction, a N force-change threshold triggered a stop request, with a s stop latency. After s of simulated contact, the force was N with the guardrail and N without it.
CdS quantum-dot synthesis
Experimental design and plate layout
The operator provided the available stock solutions, source-plate layout, reaction plate and tip inventory. SBL used these inputs and the OT-2 configuration to design a 96-well screen of CdS precursor mixtures. The screen varied the MPA:Cd molar ratio (0, 1, 1.5 and 2) and thiourea:Cd molar ratio (0.75, 1.5 and 3) across the twelve columns (Figure S4 and Table S2). Rows A–B, C–D, E–F and G–H received buffer stocks at pH 8.5, 9.0, 9.5 and 10.0, respectively, giving two wells per condition.
Figure S4. Reaction-plate design for the CdS screen. Columns represent combinations of MPA:Cd and thiourea:Cd molar ratios. Each row pair received a buffer stock at the indicated pH, giving two wells per condition. Colors distinguish the four MPA:Cd ratios.
| Column | MPA:Cd | Thiourea:Cd | Water | MPA | Thiourea |
| 1 | 0 | 0.75 | 200 | 0 | 30 |
| 2 | 0 | 1.5 | 170 | 0 | 60 |
| 3 | 0 | 3 | 110 | 0 | 120 |
| 4 | 1 | 0.75 | 160 | 40 | 30 |
| 5 | 1 | 1.5 | 130 | 40 | 60 |
| 6 | 1 | 3 | 70 | 40 | 120 |
| 7 | 1.5 | 0.75 | 140 | 60 | 30 |
| 8 | 1.5 | 1.5 | 110 | 60 | 60 |
| 9 | 1.5 | 3 | 50 | 60 | 120 |
| 10 | 2 | 0.75 | 120 | 80 | 30 |
| 11 | 2 | 1.5 | 90 | 80 | 60 |
| 12 | 2 | 3 | 30 | 80 | 120 |
Table S2. Reaction-plate design and commanded volumes per well. MPA and thiourea ratios are relative to Cd. Each column was repeated at four buffer-stock pH values, with two wells per condition. Volumes are in .
The source plate contained 1 M ammonium chloride/ammonia buffer in column 1, 10 mM Cd(NO) in column 2, 10 mM MPA in column 3, 10 mM thiourea in column 4 and deionized water in column 5. The three 10 mM stocks were used directly. Each reaction well received 90 L buffer and 40 L cadmium stock; water brought the commanded total volume to 360 L. The pH values refer to the buffer stocks.
Liquid-handling planning and execution
OT-2 operation used API, built on code from the OT2Demo repository (https://github.com/clascclon/OT2Demo). SBL inspected the OT-2 deck, pipette and available tips through the device interface and generated a Python protocol for the screen. The plan used the left-mounted eight-channel P300 pipette, a deep-well source plate in slot 5, the reaction plate on a 5 mm holder in slot 2, and tip racks in slots 10 and 11. The plate and holder heights were incorporated into the labware geometry for determining the appropriate depth for aspiration and mixing. The protocol was simulated offline and checked against the available water volume and tip inventory before execution.
The pipette first distributed water, then buffer, cadmium stock and MPA stock. The buffer, cadmium, MPA and thiourea stocks were each mixed before their first transfer. Thiourea was added last, using fresh tips for each reaction column, followed by five mixing cycles of 150 L in the reaction wells. After each transfer, the pipette blew out above the destination well before moving away. Aspiration, dispensing and blowout rates were set to 90 L s. The completed run contained 236 acknowledged device commands and used 128 tips.
Plate transfer, sealing and reader loading
After liquid handling, SBL coordinated 19 robotic stages to transfer and seal the reaction plate and load it into the Cytation 5. The arm carried the plate from the OT-2 to the PlateLoc, then transferred the sealed plate to an intermediate holder for re-grasping along its short side before placing it in the reader. The digital twin was used for motion planning, and visual and sensory feedback guided grasping and placement. The plate was registered with the reader for optical measurements.
Incubation and optical measurements
The Cytation 5 acquired a baseline at before heating the plate to . Nine heated time points were acquired at hourly intervals from 0 to 8 h. At each timepoint, fluorescence emission was measured from 450 to 680 nm in 10 nm steps with 360 nm excitation and a 7.0 mm focal height, followed by absorbance at 350, 400 and 450 nm. The first heated measurement began after the reader reached the temperature setpoint.
Optical data and analysis
The reader records contain measurements for all 96 wells at the ambient baseline and the nine heated time points. The plate spectra display absorbance in absorbance units and fluorescence in relative fluorescence units, using reader values without blank subtraction, normalization, smoothing, or interpolation. The ambient fluorescence spectrum of well C2 was excluded from display scales and plate summaries; its recorded values remain available.
HPLC drawer manipulation and plate loading
Instrument interface
The sample-loading drawer couples a spring-loaded release mechanism to a sliding assembly carrying the plate holders. Before the manipulation trials, we fitted the drawer front with a plastic fixture carrying two magnets to provide a magnetic interface with the gripper. The curved inner surface of the front panel constrained where a gripper finger could bear against it during mechanical pulling. The nominal CAD dimensions of the drawer fixture are mm; its two magnet pockets and those on the finger inserts each have a mm opening and a depth of mm (Figure S5).

Figure S5. Magnetic fixtures for HPLC manipulation. A: Drawer-side fixture. B: Left and right finger inserts. Dimensions are nominal values in millimetres and exclude mounting tape and magnets.
Development of the manipulation strategy
Development of the manipulation strategy
The digital twin represented the drawer front, magnetic counterparts and plate holders as a moving assembly, preserving their relative arrangement as the drawer moved. In the seated configuration shown in Figure S6, the nominal plate-holder CAD clearances are approximately 0.39 mm and 0.19 mm per side along the long and short axes, respectively. These clearances require submillimetre positioning during placement. Gripper release was also constrained by the surrounding drawer geometry: full opening produced finger-panel interference in the digital twin, therefore requiring partial opening for release and vertical withdrawal (Figure S6).
Figure S6. Geometric constraints of HPLC manipulation. a Drawer configurations with finger engagement behind the front panel to open the drawer along the slide axis. The inset shows approach and inward pressing to open the drawer. b Base view of the seated plate with nominal plate-holder CAD clearances. c Full gripper opening produces finger-panel interference (red), whereas 1 mm outward travel per finger permits release and withdrawal.
The initial strategy used magnetic attachment to pull the drawer outward after pressing its release mechanism (8 mm inward). During physical trials, the magnetic connection separated before the drawer fully opened, and recorded force/torque signals showed unloading during outward motion. These observations motivated a transition to mechanical engagement for the remaining pull. The agent explored candidate gripper orientations, insertion positions and pulling motions in the digital twin, refining the strategy and model using physical observations and sensor readings. In the selected strategy, the gripper withdrew from the magnetic interface, reoriented with its fingers pointing downward, and inserted one finger behind the front panel while the other remained outside. The contact location and insertion depth were constrained by the panel dimensions, the support beneath it and the clearance required by the complete gripper.
Registration and drawer-state updates
AprilTag (https://github.com/AprilRobotics/apriltag) observations provided a spatial reference for registering the HPLC instrument in the laboratory’s world frame (Figure S7). During development, observations of the closed and open drawer, together with operator-supplied measurements, were used to refine its position, panel orientation and opening geometry. Comparing these configurations helped estimate drawer displacement and update the holder positions in the twin.

Figure S7. AprilTag-based registration of the HPLC digital twin. Captured image with AprilTags on the instrument and the corresponding digital twin rendering overlaid at 50% opacity. Solid orange and dashed cyan outlines denote detected and reprojected tag boundaries, respectively.
For trials with recorded unloading events, force/torque changes and synchronized robot joint reports also informed estimates of the release state. These estimates depended on the assumption that the drawer followed the tool while engaged and stopped after detachment during slow withdrawal. State updates were made at the available observation or review stages.
Operation sequence and force/torque references
The resulting workflow combined drawer opening, plate transfer, constrained release and drawer closure (Table S3). Force/torque signals and robot joint reports were recorded together and compared by operation phase. The traces capture changes during pressing, pulling, disengagement, plate placement, and closure, providing empirical references for automatically detecting departures from previously observed responses.
| Phase | Action and relevant state |
| Release drawer | Approach the front panel, press inward and withdraw from compression. |
| Engage and pull | Disengage the magnetic interface, reorient, insert one finger behind the panel and pull along the drawer axis. |
| Transfer plate | Leave the drawer, grasp the deep-well plate on the liquid handler and transfer it above the holder at the open drawer position. |
| Place and release | Lower the plate into the holder, partially open the gripper and withdraw upward within the available clearance. |
| Close the drawer | Reposition the empty gripper, push the drawer inward, press the release mechanism and retreat. |
Table S3. Sequence of HPLC drawer manipulation and plate loading.
RoboChem-Flex flowchemistry
Information access
The agent was provided with the RoboChem-Flex publication and supporting information [22], CAD, photographs and videos, four camera feeds, and low-level device API. The supporting information details the 28 fluidic channels that track routes, syringe inventories, and liquid/gas segments. The existing orchestration software, deck map and high-level experimental workflows were withheld. The API exposed motion, pumping, valves, illumination and instrument observations. Specified limits were 1 mL min for pumps and 1000 mm min for gantries. Human operator supplied depth references, priming and source-inventory guidance.
Reconstruction and commissioning
Camera comparisons guided scene refinement. Eighteen reactor-span measurements covered 3.80–4.10 mL, consistent with the initial 3.90–4.00 mL representation. Iterative simulation and execution with camera checks were used to locate the visible tip, where the needle descended to mm. A nominal 650 L draw produced a 638–642 L optical packet width.
NMR commissioning included acquisition-parameter selection and scan-guided slug positioning. The agent settled on approximately 3.3 s acquisition time per scan. It alternated rapid two-scan spectra with small pumping increments to locate the slug in the flow cell, then held it stationary for a 64-scan acquisition.
Calibrated transport prediction
Tests used acetonitrile with the lamp off, the NMR disconnected and the outlet directed to waste. H, I and O denote the handler outlet, reactor inlet and reactor outlet sensors. Routes were either liquid-primed or nitrogen-cleared. Phase changes were bracketed by adjacent sensor readings, using request-time and pump-counter records or buffered acquisition at approximately 9 Hz per channel.
Calibration used Q1/P2/P3/P4/P8, with 16 H–I and 20 I–O boundary events. Event displacement midpoints were averaged within each run, then run means were weighted equally. This gave L and L. A boundary observed at H at time reaches sensor X at the first ,
where is pump-motor volume per unit time. At 0.78 mL min, the predicted reactor transit is 301.735 s. Pump holds add to transit time; the model assumes no coast during a stop. Recipe forecasts used a preparation template; H-conditioned forecasts used the packet measured in the current run. Both were saved before the relevant execution or first predicted inlet arrival, respectively.
Transport comparisons. Table S4 details specific packets investigated in this comparison. The transit comparison included sixteen calibration transits and twelve held-out transits from P9/P13/T2/T4. Without refitting, all twelve held-out midpoints were within 1.4 s of prediction. All eight P9/P13 brackets lay within the frozen forecast envelopes; P9 included its measured hold under the prespecified rule. Multiple boundaries within a packet are correlated observations.
| Run | Test | Role and result |
| Q1 | 0.26; liquid-filled route | Calibration; initial rate-transfer envelope. |
| P2 | 0.78; liquid-filled route | Calibration; nominal gas draws 50/75 . |
| P3, P4 | 0.78; larger rear/front gas draw | Calibration; respectively 50/150 and 150/75 gas draws. |
| P8 | 0.39; liquid-filled route | Calibration at an intermediate rate. |
| P9 | 0.78 with a measured 50–53 s stop | Held out; all four transit brackets inside the frozen envelope with the measured-hold rule. |
| P13 | 0.39; 325 liquid packet | Held out; all four transit brackets inside the frozen envelope. |
| T2, T4 | 0.78; baseline/halved injection rate | Later held-out tests; four core-boundary transits of 302.22–302.47 s versus 301.735 s predicted. |
| P6, P7 | 0.78/0.26; nitrogen-cleared route | Diagnostic bulk selection: short pulses flagged retrospectively; 13 of 14 transit brackets inside, all 14 overlapping the earlier envelopes. |
| T1, T3 | No gas caps / doubled gas-draw rate | Preparation controls; T1 merged liquid with chase, while T3 gave a 643.6 core at H. Not part of the twelve held-out transit points. |
| R1 | 7 mL at 0.78; continuous flow | Later comparison: core-front/tail transits 302.40/302.59 s versus 301.74 s predicted; outlet window 49.50 s versus 49.32 s. |
| S1 | 3.5 mL at 0.39, then 3.5 mL at 0.78 | Displacement comparison: nine associated outlet boundaries agreed with the upstream-conditioned prediction within 12 . |
| S2 | 6 mL at 0.78; 45 s hold; 1 mL at 0.78 | Completed prespecified test; fourteen downstream boundaries within 19.2 of the H-conditioned forecast. |
| S3 | 6 mL at 0.78, then 1 mL at 0.26 | Displacement comparison: nine retrospectively associated outlet boundaries agreed with prediction within 32.3 . |
Table S4. Table S4 Transport experiments and their evidential roles. Rates are in ; packet volumes and gas draws are commanded values. Unless stated otherwise, the nominal liquid packet was 650 . Calibration rows refer to the final fit, even where an earlier forecast preceded the run.
The main figure compares continuous flow (R1) with a pump pause (S2), using nominal 50/650/50 L gas/liquid/gas packets and 7 mL main-pump displacement. R1 gave core-front/tail transits of 302.40/302.59 s against 301.74 s predicted, and an outlet liquid window of 49.50 s against 49.32 s. With a measured 45.20 s hold, S2 gave an outlet window of 94.48 s against 94.60 s predicted from H. All fourteen associated downstream boundaries were within 19.2 L of prediction, meeting the chosen 75 L tolerance. Prediction and observation share the saved pump-start origin; no outlet-time alignment was fitted. S1/S3 displacement comparisons gave outlet residuals within 12/32.3 L; S1 used its measured split packet and S3 used retrospective boundary association.
Hydraulic model and choice of transport predictor
The agent also built a one-dimensional hydraulic model with gas compression and flow resistance. Both models used measured H boundaries and recorded pump execution. They were calibrated on T2/T4 and evaluated without refitting on R1/S2, with two packets and eight core-boundary observations per group. The hydraulic comparison represented the displacement-derived volumes as horizontal 0.8 mm-bore tubes and fitted two additional sensor offsets. Nominal assumptions were 0.37 mPa s liquid viscosity, 100 kPa initial gas pressure, zero surface tension and a 0.15 mL outlet extension.
Both models gave a mean absolute boundary error of about 1.81 L on the evaluation packets. Their difference (errors in pump-motor coordinates) was much smaller than the observation brackets. The displacement model (Equation (1)) was retained because it supplied sufficient working predictions with fewer fitted quantities and no gas-pressure initialization.
The hydraulic solver uses cumulative tube volume,
where is the internal radius. For gas cap , interfaces enclose volume at absolute pressure . Adjacent liquid flows give
The pump sets . Each incompressible liquid plug obeys
with viscosity , density and elevation change . The optional capillary closure is , with surface tension and fixed contact angle . An exposed gas region of volume vents according to
where is the gas-region Poiseuille resistance. Unlike displacement transport, this model permits pressure relaxation and liquid motion during a pump hold. It remains available for exploring tube dimensions, reactor volumes and pump schedules. Predictions depend on assumed geometry and fluid properties; films, breakup, contact-angle hysteresis, multidimensional mixing and irradiation fields are not resolved.
Photochemistry and spectral processing
All chemicals were sourced from Merck KGaA and used without further purification. Each nominal 650 L packet contained 48.75 mol benzylidenemalononitrile (75 mM), neat THF and TBADT in acetonitrile. Stock draws were 24.375 L of 2.0 M substrate, 4.0625 L THF for equivalents and 28.676 L of 17 mM catalyst for mol%, with acetonitrile to volume. Seven reciprocating 175 L sampler strokes preceded gas-capped injection. Flow rates of 0.26/0.39/0.78 mL min gave nominal 15/10/5 min residence times in the approximately 3.9 mL reactor under 370 nm illumination.
After scan-guided positioning, 64-scan H spectra were acquired at 61.8 MHz with WET acetonitrile suppression and C decoupling (Figure S8). Acquisition time was 3.3 s, repetition time 20 s, and receiver gain 25; 8192 points were zero-filled to 32768. Each acquisition lasted approximately 301 s. Table S5 gives the conditions and spectral ratios. O8 failed preparation and supplied no 64-scan spectrum. The fragmented O9 packet was excluded before scoring. X was a separate condition check; a substrate reference used 650 L of 2.0 M stock through the dark reactor. The score, specified before O1, was the signed product-window integral divided by the aromatic-window integral after linear baseline correction. Native-scale windows were 3.0–4.0 and 5.7–7.1 ppm; baseline flanks were 2.85–2.98/4.10–4.30 and 5.30–5.55/7.30–7.55 ppm, respectively. Windows followed each run’s acetonitrile offset relative to 0.802 ppm. Trapezoidal integration retained negative values and reproduced archived scores within 0.2%.
Figure S8. Broader NMR spectra of O1–O7. Instrument-processed exports are display-aligned on a common intensity scale, without additional baseline correction, per-trace normalisation or vertical clipping. Gold/teal indicate 5/15 equivalents of THF. Grey/blue bands mark aromatic/product integration regions.
| Run | THF | TBADT | Lamp | Ratio | |
| Initial | 5 | 1 | 10 | 50 | 0.3068 |
| O1 | 5 | 1 | 10 | 100 | 0.3109 |
| O2 | 15 | 2 | 15 | 100 | 0.3236 |
| O3 | 15 | 1 | 10 | 100 | 0.3438 |
| O4 | 5 | 2 | 15 | 100 | 0.2987 |
| O5 | 5 | 2 | 5 | 100 | 0.3143 |
| O6 | 15 | 1 | 15 | 100 | 0.3604 |
| O7 | 15 | 1 | 5 | 100 | 0.3557 |
| O8 | 15 | 2 | 5 | 100 | — |
| O9 | 15 | 1 | 15 | 100 | 0.3052 |
| X | 11 | 1.5 | 5 | 75 | 0.3318 |
Table S5. Reaction conditions and corrected spectral ratios. THF is in equivalents, TBADT in mol%, in minutes and lamp setting in percent. Each condition was run once. O1–O7 form the main comparison.
LumaStir holder fabrication and operation
Holder design and fabrication
The in-house LumaStir device accommodates two rows of three vials for illumination and stirring. The print mesh for the front row had three circular top openings, each 18.0 mm in diameter, on 40.0 mm centers and a front window below each opening for viewing the sample. The mesh measured mm. The front slot of the assembly model has an envelope of mm. Centering the new holder in this envelope gave a nominal clearance of 0.25 mm on each side along both horizontal axes.
The holder was sliced for a Bambu Lab A1 with a 0.4 mm nozzle and printed in poly(lactic acid) on a textured polyetherimide plate. The submitted settings specified a 0.2 mm layer height and commanded nozzle and bed temperatures of 220 and 65 °C, respectively. The digital twin combined the print layout with the programmed final bed position of mm to estimate the holder pose after printing, using a nominal mapping between the printer and twin bed coordinates. The LumaStir folder in Data S1 contains the holder mesh (LumaStir-B.stl), slicer project (LumaStir-B-project.3mf), and print package (LumaStir-B-print.3mf). The accompanying wrist camera files include color and aligned depth images of the holder on the print bed before pickup and a color image of the holder above the front slot before the lateral adjustment.
Robot pickup and installation
With the UR5e stationary, its wrist-mounted D435i camera recorded the holder on the print bed. Lines fitted to the four top edges in the color image located the corners, and a plane fitted to the aligned depth measurements described the top surface. Intersecting camera rays through the corners with this plane gave four points that were mapped into the twin to estimate the holder center and long-axis orientation. Relative to the nominal model, the estimated center displacement was () mm in the twin world axes and the long-axis yaw difference was modulo . This single-view estimate depends on the depth measurements, hand-eye calibration and robot kinematics. The UR5e picked up the holder using the updated grasp target and transported it to LumaStir.
The UR5e positioned the holder above the slot, where a second wrist-camera observation guided a small lateral adjustment before a force-guided insertion search. The UR5e released the holder and pressed it into place with its closed gripper. The reported peak force change relative to the start of the press was 4.13 N. The UR5e then withdrew and reopened its gripper. Images showed that the holder protruded less after pressing. The controller recorded tool poses in the UR5e base frame. Estimating the seated holder position from these poses required an assumed grasp offset and the support height from the assembly model.
Vial handoff and device operation
The recorded handoff began with the UR3e holding a prepared vial. The twin plan used a nominal pose for the vial relative to the gripper. At the planned handoff pose, the UR5e received the vial from the UR3e. The UR3e then retracted, and the UR5e carried the vial to LumaStir, placed it in the rightmost holder opening and withdrew. For the rightmost slot in the holder, the controller received a brightness setting of 100 for light channel 2 and a speed setting of 100 for motor channel 2, with a motor run time of 60 s.
References
- [4]Connor W Coley, Dale A Thomas, Justin A M Lummiss, Jonathan N Jaworski, Christopher P Breen, Victor Schultz, Travis Hart, Joshua S Fishman, Luke Rogers, Hanyu Gao, Robert W. Hicklin, Pieter P. Plehiers, Joshua Byington, John S. Piotti, William H. Green, A. John Hart, Timothy F. Jamison, and Klavs F. Jensen. A Robotic Platform for Flow Synthesis of Organic Compounds Informed by AI Planning. Science, 365(6453):eaax1566, 2019. doi:10.1126/science.aax1566.DOI
- [6]Benjamin Burger, Phillip M Maffettone, Vladimir V Gusev, Catherine M Aitchison, Yang Bai, Xiaoyan Wang, Xiaobo Li, Ben M Alston, Buyi Li, Rob Clowes, Nicola Rankin, Brandon Harris, Reiner Sebastian Sprick, and Andrew I. Cooper. A Mobile Robotic Chemist. Nature, 583(7815):237–241, 2020. doi:10.1038/s41586-020-2442-2.DOI
- [11]Aidan Slattery, Zhenghui Wen, Pauline Tenblad, Jesús Sanjosé-Orduna, Diego Pintossi, Tim den Hartog, and Timothy Noël. Automated self-optimization, intensification, and scale-up of photocatalysis in flow. Science, 383(6681):eadj1817, 2024. doi:10.1126/science.adj1817.DOI
- [12]Jiaru Bai, Sebastian Mosbach, Connor J. Taylor, Dogancan Karan, Kok Foong Lee, Simon D. Rihm, Jethro Akroyd, Alexei A. Lapkin, and Markus Kraft. A Dynamic Knowledge Graph Approach to Distributed Self-Driving Laboratories. Nat. Commun., 15:462, 2024. doi:10.1038/s41467-023-44599-9.DOI
- [16]Richard B. Canty and Milad Abolhasani. The past, present and future of self-driving laboratories. Nature Reviews Chemistry, 10(8):523–537, 2026. ISSN 2397-3358. doi:10.1038/s41570-026-00847-2.DOI
- [17]Jiaru Bai, Liwei Cao, Sebastian Mosbach, Jethro Akroyd, Alexei A Lapkin, and Markus Kraft. From Platform to Knowledge Graph: Evolution of Laboratory Automation. JACS Au, 2(2):292–309, 2022. doi:10.1021/jacsau.1c00438.DOI
- [18]Fuzhan Rahmanian, Jackson Flowers, Dan Guevarra, Matthias Richter, Maximilian Fichtner, Phillip Donnelly, John M Gregoire, and Helge S Stein. Enabling Modular Autonomous Feedback-Loops in Materials Science through Hierarchical Experimental Laboratory Automation and Orchestration. Adv. Mater. Interfaces, 9(8):2101987, 2022. doi:10.1002/admi.202101987.DOI
- [19]Benjamin P. MacLeod, Fraser G. L. Parlane, and Curtis P. Berlinguette. How to build an effective self-driving laboratory. MRS Bulletin, 48(2):173–178, 2023. doi:10.1557/s43577-023-00476-w.DOI
- [20]Melodie Christensen, Lars PE Yunker, Parisa Shiri, Tara Zepel, Paloma L Prieto, Shad Grunert, Finn Bork, and Jason E Hein. Automation isn’t Automatic. Chem. Sci., 12(47):15473–15490, 2021. doi:10.1039/D1SC04588A.DOI
- [21]Anne-Catherine Bédard, Andrea Adamo, Kosi C Aroh, M Grace Russell, Aaron A Bedermann, Jeremy Torosian, Brian Yue, Klavs F Jensen, and Timothy F Jamison. Reconfigurable System for Automated Optimization of Diverse Chemical Reactions. Science, 361(6408):1220–1225, 2018. doi:10.1126/science.a at0650.DOI
- [22]Simone Pilon, Elia Savino, Oliver M. Bayley, Michael Vanzella, Miguel Claros, Petros Siasi aridis, Junsong Liu, Florian Lukas, Matteo Damian, Vasilis Tseliou, Niccolò Intini, Aidan Slattery, Jesus SanJosé-Orduna, Tim den Hartog, Ron A. H. Peters, Andrea F. G. Gargano, Francesco G. Mutti, and Timothy Noël. A flexible and affordable self-driving laboratory for automated reaction optimization. Nature Synthesis, 2026. doi:10.1038/s44160-026-01053-0.DOI
- [23]Yunheng Zou, Austin H. Cheng, Abdulrahman Aldossary, Jiaru Bai, Shi Xuan Leong, Jorge Arturo Campos-Gonzalez-Angulo, Changhyeok Choi, Cher Tian Ser, Gary Tom, Andrew Wang, Zijian Zhang, Ilya Yakavets, Han Hao, Chris Crebolder, Varinia Bernales, and Alán Aspuru-Guzik. El Agente: An autonomous agent for quantum chemistry. Matter, 8(7):102263, 2025. doi:10.1016/j.matt.2025.102263.DOI
- [24]Naruki Yoshikawa, Marta Skreta, Kourosh Darvish, Sebastian Arellano-Rubach, Zhi Ji, Lasse Bjørn Kristensen, Andrew Zou Li, Yuchi Zhao, Haoping Xu, Artur Kuramshin, Alán Aspuru-Guzik, Florian Shkurti, and Animesh Garg. Large language models for chemistry robotics. Autonomous Robots, 2023. doi:10.1007/s10514-023-10136-2.DOI
- [25]Sebastian Pagel, Michael Jirasek, and Leroy Cronin. Verification and execution of the scientific literature via computation augmented by large language models. Communications Chemistry, 9:191, 2026. doi:10.1038/s42004-026-01993-w.DOI
- [26]Daniil A. Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes. Autonomous Chemical Research with Large Language Models. Nature, 624(7992):570–578, 2023. doi:10.1038/s41586-023-06792-0.DOI
- [27]Andres M. Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D. White, and Philippe Schwaller. Augmenting Large Language Models with Chemistry Tools. Nat. Mach. Intell., 6(5):525–535, 2024. doi:10.1038/s42256-024-00832-8.
- [28]Keyan Ding, Jing Yu, Junjie Huang, Yuchen Yang, Qiang Zhang, and Huajun Chen. SciToolAgent: A knowledge-graph-driven scientific agent for multitool integration. Nat. Comp. Sci., 5(10):962–972, 2025. doi:10.1038/s43588-025-00849-y.DOI
- [29]Jiaru Bai, Abdulrahman Aldossary, Thomas Swanick, Marcel Müller, Yeonghun Kang, Changhyeok Choi, Naruki Yoshikawa, Zijian Zhang, Jin Won Lee, Tsz Wai Ko, Aiwei Yin, Mohammad Ghazi Vakili, Chris Crebolder, Varinia Bernales, and Alán Aspuru-Guzik. El agente gráfico: A semantic execution runtime for scientific agents. arXiv Preprint, 2026. doi:10.48550/arXiv.2602.17902.DOI
- [30]Marcel Müller, Jiaru Bai, Willi Gottstein, Abhijoy Mandal, Mohammad Nazeri, Elia Savino, Yanlin Fang, Sujoy Das, Sergio Pablo García Carrillo, Yeonghun Kang, Juan B. Pérez-Sánchez, Simone Pilon, Martin Fitzner, Timothy Noël, Frank Gu, Varinia Bernales, and Alán Aspuru-Guzik. La agente Óptima: Towards agentic self-driving laboratories. arXiv Preprint, 2026. doi:10.48550/arXiv.2609.04564.DOI
- [31]Kourosh Darvish, Arjun Sohal, Abhijoy Mandal, Hatem Fakhruldeen, Nikola Radulov, Zhengxue Zhou, Satheeshkumar Veeramani, Joshua Choi, Sijie Han, Brayden Zhang, Jeeyeoun Chae, Alex Wright, Yijie Wang, Hossein Darvish, Yuchi Zhao, Gary Tom, Han Hao, Miroslav Bogdanovic, Gabriella Pizzuto, Andrew I. Cooper, Alán Aspuru Guzik, Florian Shkurti, and Animesh Garg. MATTERIX: toward a digital twin for robotics-assisted chemistry laboratory automation. Nat. Comput. Sci., 6:67–82, 2026. doi:10.1038/s43588-025-00924-4.DOI
- [32]S. D. Rihm, J. Bai, A. Kondinski, S. Mosbach, J. Akroyd, and M. Kraft. Transforming Research Laboratories with Connected Digital Twins. Nexus, 1(1):100004, 2024. doi:10.1016/j.ynex.2024.100004.DOI
- [33]Kourosh Darvish, Marta Skreta, Yuchi Zhao, Naruki Yoshikawa, Sagnik Som, Miroslav Bogdanovic, Yang Cao, Han Hao, Haoping Xu, Alán Aspuru-Guzik, Animesh Garg, and Florian Shkurti. ORGANA: A Robotic Assistant for Automated Chemistry Experimentation and Characterization. Matter, 8(2):101897, 2024. doi:10.1016/j.matt.2024.10.015.DOI
- [34]Tianwei Dai, Sriram Vijayakrishnan, Filip T. Szczypiński, Jean-François Ayme, Ehsan Simaei, Thomas Fellowes, Rob Clowes, Lyubomir Kotopanov, Caitlin E. Shields, Zhengxue Zhou, John W. Ward, and Andrew I. Cooper. Autonomous mobile robots for exploratory synthetic chemistry. Nature, 635(8040):890–897, 2024. ISSN 1476-4687. doi:10.1038/s41586-024-08173-7.DOI
- [35]Naruki Yoshikawa, Andrew Zou Li, Kourosh Darvish, Yuchi Zhao, Haoping Xu, Artur Kuramshin, Alán Aspuru-Guzik, Animesh Garg, and Florian Shkurti. Chemistry lab automation via constrained task and motion planning. arXiv Preprint, 2023. doi:10.48550/arXiv.2212.09672.DOI
- [36]Louis Longley, Francisco Munguia-Galeano, Yushu Han, Rob Clowes, Sriram Vijayakrishnan, Adam Edwards, Gabriella Pizzuto, Hatem Fakhruldeen, and Andrew Cooper. Robinhood: a robotic chemist in a fume hood. Digital Discovery, 5(6):2438–2447, 2026. ISSN 2635-098X. doi:10.1039/d6dd00007j.DOI
- [37]Yuchi Zhao, Miroslav Bogdanovic, Chengyuan Luo, Steven Tohme, Kourosh Darvish, Alan Aspuru-Guzik, Florian Shkurti, and Animesh Garg. Anyplace: Learning generalizable object placement for robot manipulation. In Joseph Lim, Shuran Song, and Hae-Won Park, editors, Proceedings of The 9th Conference on Robot Learning, volume 305 of Proceedings of Machine Learning Research, pages 4038–4057. PMLR, 27–30 Sep 2025. https://proceedings.mlr.press/v305/zhao25b.html.arxiv.org/abs/2502.04531
- [38]Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthan Gopalakrishnan, Karol Hausman, Alex Herzog, Daniel Ho, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Eric Jang, Rosario Jauregui Ruano, Kyle Jeffrey, Sally Jesmonth, Nikhil J Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Kuang-Huei Lee, Sergey Levine, Yao Lu, Linda Luu, Carolina Parada, Peter Pastor, Jornell Quiambao, Kanishka Rao, Jarek Rettinghouse, Diego Reyes, Pierre Sermanet, Nicolas Sievers, Clayton Tan, Alexander Toshev, Vincent Vanhoucke, Fei Xia, Ted Xiao, Peng Xu, Shichun Xu, Mengyuan Yan, and Andy Zeng. Do as i can, not as i say: Grounding language in robotic affordances. arXiv Preprint, 2022. doi:10.48550/arXiv.2204.01691.
- [39]Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. arXiv Preprint, 2022. doi:10.48550/arXiv.2209.07753.arxiv.org/abs/2209.07753
- [40]Phillip Isola. Robot-use agents. Blog post, September 2026. https://web.mit.edu/phillipi/www/writing/robot-use-agents.html. Published 7 September 2026; accessed 16 October 2026.
- [41]Wenbo Zhang, Kaixuan Wang, Yutao Ouyang, Xiaoyu Huang, Liyang Li, Kailun Su, Weiyang Jin, Wenhao Chai, Haotian Liang, Zhiyang Dou, Yue Chen, and Tianxing Chen. An unexpected robot policy: Early evaluations of gpt-6 astra on robodojo and beyond. arXiv Preprint, 2026. doi:10.48550/arXiv.2609.24170.DOI
- [42]Galbot Team, Xuchuan Chen, Xiaoqian Cheng, Yu Deng, Lihe Ding, Shaocong Dong, Xiangjun Gao, Haozhe Jia, Zekai Li, Zhoujian Li, Yunrui Lian, Sikai Liang, Chenghuai Lin, Dairu Liu, Jiahง Liu, Qingtao Liu, Yuxuan Ma, Zekun Qi, Jiayi Su, He Wang, Ruocheng Xu, Tianyu Xu, Xudong Xu, Zhe Xu, Mi Yan, Siming Yan, Li Yi, Ruixi Yu, Jinlu Zhang, Yintianrun Zhang, Zhikai Zhang, Zhizheng Zhang, Yixin Zheng, and Weiyi Zhu. Systematically exploring the capabilities of gpt-6 astra as embodied policies. arXiv Preprint, 2026. doi:10.48550/arXiv.2609.38537.DOI
- [43]Jenai Xuning Yang, Ashna Khetan, and Shun Zhang. Astronomy: An empirical study of GPT-6 Astra as a native robot policy. NVIDIA Seattle Robotics Lab technical report, 2026. https://research.nvidia.com/labs/srl/projects/astronomy/. Accessed 16 October 2026.
- [44]Sida He, Lingxi Xie, Yunning Cao, Pengfei Chen, Kaiwen Duan, Jiannan Ge, Xinyue Huo, Jiacheng Shao, and Qi Tian. Robot manipulation with gpt-6-astra: Body knowledge, experience reuse, emergent skills, and sim2real transfer. arXiv Preprint, 2026. doi:10.48550/arXiv.2609.31770.DOI
- [45]Gio Huh, Cayden Gu, Takara E. Truong, C. Karen Liu, and Guy Tevet. HomeBody: A humanoid that explores, remembers, and acts on its own. Project technical report, 2026. https://tml.stanford.edu/homebody/. Accessed 16 October 2026.
- [46]Achu Menon, Sravanthi Machcha, Sabrina Zou, Tzu Kit Chan, and Jay Chooi. GPT-6 Astra on robotic manipulation. Robocurve benchmark report, September 2026. https://openai.robocurve.org/gpt-6-astra/. Published 4 September 2026; accessed 16 October 2026.
- [47]Andrew I. Cooper, Patrick Courtney, Kourosh Darvish, Moritz Eckhoff, Hatem Fakhruldeen, Andrea Gabrielli, Animesh Garg, Sami Haddadin, Kanako Harada, Jason Hein, Maria Hübner, Dennis Knobbe, Gabriella Pizzuto, Florian Shkurti, Ruja Shrestha, Kerstin Thurow, Rafael Vescovi, Birgit Vogel-Heuser, Ádam Wolf, Naruki Yoshikawa, Yan Zeng, Zhengxue Zhou, and Henning Zwirnmann. Accelerating discovery in natural science laboratories with ai and robotics: Perspectives and challenges. Science Robotics, 10(106), 2025. doi:10.1126/scirobotics.adv7932.DOI
- [48]Balakumar Sundaralingam, Siva Kumar Sastry Hari, Adam Fishman, Caelan Garrett, Karl Van Wyk, Valts Blukis, Alexander Millane, Helen Oleynikova, Ankur Handa, Fabio Ramos, Nathan Ratliff, and Dieter Fox. Curobo: Parallelized collision-free robot motion generation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8112–8119, 2023. doi:10.1109/ICRA48891.2023.10160765.DOI
- [49]Shi Xuan Leong, Caleb E Griesbach, Rui Zhang, Kourosh Darvish, Yuchi Zhao, Abhijoy Mandal, Yunheng Zou, Han Hao, Varinia Bernales, and Alán Aspuru-Guzik. Steering towards safe self-driving laboratories. Nat. Rev. Chem., 9(10):707–722, 2025. doi:10.1038/s41570-025-00747-x.DOI
- [50]Yeonghun Kang, Shi Xuan Leong, Hyun Suk Park, Jiaru Bai, Sven Papidocha, Yuchi Zhao, Rui Zhang, Magali Gimeno, Stanley Lo, Eric S. Isbrandt, Han Hao, Yang Cao, Felix Katzenburg, Fangshi Du, Alán Aspuru-Guzik, and Varinia Bernales. El agente seguro: An agent for chemical safety. 2026. doi:10.26434/chemrxiv.15004378/v2.DOI
- [51]Linjian Chen, Xiaobo Li, Quan Lin, and Jun Jiang. Self-driving laboratories need an autonomy safety harness. Nat. Synth., 2026. doi:10.1038/s44160-026-01120-6.DOI
- [52]Abdelhay Abouliaich, Denis Billaud, Mouhammad Abyan, Lavinia Balan, Jean-Jacques Gaumet, Ghouti Medjadhi, Jaafar Ghanbaja, and Raphaël Schneider. One-pot noninjection route to cds quantum dots via hydrothermal synthesis. ACS Appl. Mater. Interfaces, 4(5):2561–2569, 2012. doi:10.1021/am300232z.
- [53]Naim Bel Haj Mohamed, Nassim Ben Brahim, Randa Mrad, Mohamed Haouari, Rafik Ben Chaâbane, and Michel Negreie. Use of MPA-capped CdS quantum dots for sensitive detection and quantification of Co2+ ions in aqueous solution. Analytica Chimica Acta, 1028:50–58, 2018. doi:10.1016/j.aca.2018.04.041.
- [54]Davide Ravelli, Stefano Protti, and Maurizio Fagnoni. Decatungstate anion for photocatalyzed “window ledge” reactions. Accounts of Chemical Research, 49(10):2232–2242, 2016. doi:10.1021/acs.accounts.6b00339.DOI
- [55]Gabriele Laudadio, Yuchao Deng, Klaas van der Wal, Davide Ravelli, Manuel Nuño, Maurizio Fagnoni, Duncan Guthrie, Yuhan Sun, and Timothy Noël. (Cp₃)–H functionalizations of light hydrocarbons using decatungstate photocatalysis in flow. Science, 369(6499):92–96, 2020. doi:10.1126/science.abb4688.DOI
- [56]Lulu Guo, Yingkai Sun, Xiaobo Li, Luyao Ge, Ziming Wang, Haitao Zheng, Jingyu Li, Huijuan Zhang, Bingxu Chen, Daobin Liu, Yuebo Liu, Jie Li, Xiaohui Li, Linjiang Chen, Yi Luo, and Jun Jiang. Stress-testing large language model agents in a robotic chemistry laboratory. arXiv Preprint, 2026. doi:10.48550/arXiv.2607.23045.DOI
- [57]Henning Bär, Remo Hochstrasser, and Bernd Papenfuß. SiLA: Basic Standards for Rapid Integration in Laboratory Automation. SLAS Technol., 17(2):86–95, 2012. doi:10.1177/221106821244550.DOI
- [58]Zhengxue Zhou, Satheeshkumar Veeramani, Francisco Munguia-Galeano, Hatem Fakhruldeen, and Andrew I. Cooper. Localization, inspection, and reasoning (lira) module for autonomous workflows in self-driving laboratories. Communications Chemistry, 8(1), 2025. ISSN 2399-3669. doi:10.1038/s42004-025-01770-1.DOI
- [59]Edward Sun, Sravanthi Machcha, Sabrina Zou, Tzu Kit Chan, and Jay Chooi. RoboHarm: Do frontier robot policies refuse unsafe instructions? Robocurve benchmark report, September 2026. https://robocurve.org/roboharm/. Published 18 September 2026; accessed 16 October 2026.
- [60]Alberto Ferrari and Karen E. Willcox. Digital twins in mechanical and aerospace engineering. Nat. Comput. Sci., 4:178–183, 2024. doi:10.1038/s43588-024-00613-8.DOI
- [61]Tianle Cai, Xuezhi Wang, Tengyu Ma, Xinyun Chen, and Denny Zhou. Large Language Models as Tool Makers. arXiv Preprint, 2023. doi:10.48550/arXiv.2305.17126.DOI
- [62]Xu Huang, Junwu Chen, Yuxing Fei, Zhuohan Li, Philippe Schwaller, and Gerbrand Ceder. Cascade: Cumulative agentic skill creation through autonomous development and evolution. arXiv Preprint, 2025. doi:10.48550/arXiv.2512.23880.DOI
- [63]Robert Rauschen, Mason Guy, Jason E. Hein, and Leroy Cronin. Universal Chemical Programming Language for Robotic Synthesis Repeatability. Nat. Synth., 2024. doi:10.1038/s44160-023-00473-6.DOI
- [64]Ruiyang Si, Jianxin Bi, Shunyu Yang, Rui Ni, Wenbo Huang, Qiang Wang, Shulong Jiang, Yintianrun Zhang, Feixiang Ruan, Kewei Zhang, Duomin Wang, Xuyu Li, Haiwen Feng, Zhen Dong, and Daquan Zhou. Fewer tokens, better action: Gpt-6 astra robot agents with 14% higher success rate but 65% fewer tokens. arXiv Preprint, 2026. doi:10.48550/arXiv.2610.01939.DOI
- [65]Felix Strieth-Kalthoff, Han Hao, Vandana Rathore, Joshua Derasp, Théophile Gaudin, Nicholas H. Angello, Martin Seifrid, Ekaterina Trushina, Mason Guy, Jullian Liu, Xun Tang, Masashi Mamada, Wesley Wang, Tuul Tsagaantsooj, Cyrille Lavigne, Robert Pollice, Tony C. Wu, Kazuhiro Hotta, Leticia Bodo, Shangyu Li, Mohammad Haddadnia, Agnieszka Wołos, Rafał Roszak, Cher Tian Ser, Carlota Bozal-Ginesta, Riley J. Hickman, Jenya Vestfrid, Andrés Aguilar-Granda, Elena L. Klimareva, Ralph C. Sigerson, Wenduan Hou, Daniel Gahler, Slawomir Lach, Adrian Warzybok, Oleg Borodin, Simon Rohrbach, Benjamin Sanchez-Lengeling, Chihaya Adachi, Bartosz A. Grzybowski, Leroy Cronin, Jason E. Hein, Martin D. Burke, and Alán Aspuru-Guzik. Delocalized, asynchronous, closed-loop discovery of organic laser emitters. Science, 384(6697), 2024. doi:10.1126/science.adk9227.
- [66]Chang Jie Leong, Kai Yuan Andre Low, Jose Recatala-Gomez, Pablo Quijano Velasco, Eleonore Vissol-Gaudin, Jin Da Tan, Balamurugan Ramalingam, Riko I Made, Shreyas Dinesh Pethe, Saumya Sebastian, Yee-Fun Lim, Zi Hui Jonathan Khoo, Yang Bai, Jayce Jian Wei Cheng, and Kedar Hippalgaonkar. An object-oriented framework to enable workflow evolution across materials acceleration platforms. Matter, 5(10):3124–3134, 2022. doi:10.1016/j.matt.2022.08.017.DOI
- [67]Eric Stach, Brian DeCost, A Gilad Kusne, Jason Hattrick-Simpers, Keith A Brown, Kristofer G Reyes, Joshua Schrier, Simon Billinge, Tonio Buonassisi, Ian Foster, Carla P. Gomes, John M. Gregoire, Apurva Mehta, Joseph Montoya, Elsa Olivetti, Chiwoo Park, Eli Rotenberg, Semon K. Saikin, Sylvia Smullin, Valentin Stanev, and Benji Maruyama. Autonomous Experimentation Systems for Materials Development: A Community Perspective. Matter, 4(9):2702–2726, 2021. doi:10.1016/j.matt.2021.06.036.DOI
- [68]Richard B. Canty, Jeffrey A. Bennett, Keith A. Brown, Tonio Buonassisi, Sergei V. Kalinin, John R. Kitchin, Benji Maruyama, Robert G. Moore, Joshua Schrier, Martin Seifrid, Shijing Sun, Tejs Vegge, and Milad Abolhasani. Science acceleration and accessibility with self-driving labs. Nature Communications, 16(1), 2025. ISSN 2041-1723. doi:10.1038/s41467-025-59231-1.DOI
- [69]Lauren M. Sanders, Ryan T. Scott, Jason H. Yang, Amina Ann Qutub, Hector Garcia Martin, Daniel C. Berrios, Jaden J. A. Hastings, Jon Rask, Graham Mackintosh, Adrienne L. Hoar frost, Stuart Chalk, John Kalantari, Kia Khezeli, Erik L. Antonsen, Joel Babdor, Richard Barker, Sergio E. Baranzini, Afshin Beheshti, Guillermo M. Delgado-Aparicio, Benjamin S. Glicksberg, Casey S. Greene, Melissa Haendel, Arif A. Hamid, Philip Heller, Daniel Jamieson, Katelyn J. Jarvis, Svetlana V. Komarova, Matthieu Komorowski, Prachi Kothiyal, Ashish Mahabal, Uri Manor, Christopher E. Mason, Mona Matar, George I. Mias, Jack Miller, Jerry G. Myers, Charlotte Nelson, Jonathan Oribello, Seung min Park, Patricia Parsons-Wingerter, R. K. Prabhu, Robert J. Reynolds, Amanda Saravia-Butler, Suchi Saria, Aenor Sawyer, Nitin Kumar Singh, Michael Snyder, Frank Soboczenski, Karthik Soman, Corey A. Theriot, David Van Valen, Kasthuri Venkateswaran, Liz Warren, Liz Worthey, Marinka Zitnik, and Sylvain V. Costes. Biological Research and Self-Driving Labs in Deep Space Supported by Artificial Intelligence. Nat. Mach. Intell., 5(3):208–219, 2023. doi:10.1038/s42256-023-00618-4.DOI
- [70]Qing Zhu, Yan Huang, Donglai Zhou, Luyuan Zhao, Lulu Guo, Ruyu Yang, Zixu Sun, Man Luo, Fei Zhang, Hengyu Xiao, Xinsheng Tang, Xuchun Zhang, Tao Song, Xiang Li, Baochen Chong, Junyi Zhou, Yihan Zhang, Baicheng Zhang, Jiaqi Cao, Guozhen Zhang, Song Wang, Guilin Ye, Wanjun Zhang, Haitao Zhao, Shuang Cong, Huirong Li, Li-Li Ling, Zhe Zhang, Weiwei Shang, Jun Jiang, and Yi Luo. Automated synthesis of oxygen-producing catalysts from martian meteorites by a robotic ai chemist. Nature Synthesis, 3(3):319–328, 2023. ISSN 2731-0582. doi:10.1038/s44160-023-00424-4.DOI
- [71]NVIDIA. Isaac Sim: Robotics simulation and synthetic data generation. Official software framework and documentation, 2026. https://developer.nvidia.com/isaac/sim. Accessed 28 September 2026.
- [72]Stephen H. Smallcombe, Steven L. Patt, and Paul A. Keifer. WET solvent suppression and its applications to LC NMR and high-resolution NMR spectroscopy. Journal of Magnetic Resonance, Series A, 117(2):295–303, 1995. doi:10.1006/jmra.1995.0759.DOI












