ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning

Links: Project Page

Introduction

Building general-purpose action models that transfer across tasks, objects, environments, and embodiments is a central goal of robot learning. Recent vision-language-action models (VLAs) ([9, 55, 67, 41, 34, 6] (Black et al. 2024)) have made substantial progress by transferring pretrained vision-language knowledge and scaling across heterogeneous robot demonstrations. Nevertheless, their policy learning and behavioral coverage remain largely tied to action-labeled robot trajectories, whose collection requires physical platforms, embodiment-specific interfaces, and substantial human effort. Human videos offer a complementary source of experience with far greater behavioral and environmental diversity, but do not provide directly executable robot actions.

Existing approaches seek to extract actionable supervision from human videos along two main directions. Latent action models ([76, 10, 21] (Sun et al. 2026)) infer transition variables from pairs or sequences of frames and use them for representation or policy learning. While naturally scalable to action-unlabeled videos, these inferred variables are underconstrained and may capture appearance changes or camera motion rather than controllable behavior. In contrast, fully supervised approaches recover hand poses, object interactions, or retargeted trajectories to provide explicit motion supervision ([36, 82] (Liu et al. 2026)). Such supervision provides direct physical grounding, but relies on accurate pose estimation and human-to-robot correspondence. Human-video learning therefore yields a trade-off between stronger action grounding from explicit supervision and broader behavioral coverage from unlabeled videos.

Future observations provide a natural supervisory signal to bridge this gap as every video records the consequences that follow a behavior, regardless of embodiment. Predictive world models ([2, 84, 3]) and foresight-based VLAs ([80, 12] (Sun et al. 2026; Li et al. 2026)) exploit this signal by forecasting future pixels, visual features, or trajectories, providing a direct account of what changes in the world. However, inferring the intermediate dynamics that result in these changes remains challenging, particularly from large-scale human videos without action annotations. Recent reasoning-augmented VLAs offer a useful perspective by introducing intermediate notations, such as textual plans ([79] (Huang et al. 2026)), visual guidance ([43, 25]), continuous latents (Huang et al., 2026), or coarse action trajectories ([83]), to structure action generation. This raises a natural question of whether such intermediate representations can instead be grounded in the dynamics they induce in the world, providing a shared action interface across embodiments.

Motivated by this perspective, we introduce ThinkWorld, a VLA learning framework that formulates action thoughts as Chain-of-Dynamics (CoD) and reinforces them through world imagination. CoD represents an action as a sequence of discrete motion tokens that captures how the world evolves over time, providing a common reasoning interface across human and robot behaviors. We first initialize this space with action-labeled demonstrations, grounding CoD in physically meaningful behavior. ThinkWorld then predicts the future world representation induced by each CoD and compares it with the future observation, using their alignment to reinforce action reasoning. As a result, videos even without motion annotations provide a learning signal, allowing large-scale human videos and robot trajectories to jointly refine the same CoD space. This training objective propagates through the shared VLM, enriching the representations for action prediction with knowledge of how motion transforms the world. Across simulation benchmarks and real-robot evaluations, we demonstrate that ThinkWorld enables human-to-robot skill composition and generalization to unseen scenarios.

Our contributions are as follows:

• We introduce ThinkWorld, a VLA learning framework that couples a common VLM-native action-reasoning interface for both human and robot with internal world imagination.

• We propose consequence-grounded reinforcement learning, which rewards CoD according to its consistency with observed future consequences and thereby enables action reasoning to improve from videos without action labels.

• We show that ThinkWorld benefits from scaling action-unlabeled human video and enables human-to-robot skill transfer, allowing robots to execute complete tasks demonstrated in human videos while receiving robot supervision only for the constituent atomic skills.

Related Works

Generalist Robot Policies. Vision-Language-Action (VLA) models [41, 6, 34, 71, 66] (Black et al. 2024; Chen et al. 2025) build upon advances in vision-language models (VLMs) [4, 48, 53] (Chen et al. 2026; Ye et al. 2026) to learn generalist manipulation policies that map visual observations and language instructions to executable actions.. Despite acquiring broad capabilities from large-scale robot demonstrations, end-to-end imitation remains bounded by the behaviors observed during training. Recent embodied reasoning methods introduce intermediate reasoning between perception and control [79, 33, 43, 33, 25, 83, 54] (Liu et al. 2026). ThinkAct ThinkAct (Huang et al., [33] reinforces textual and visual plans with action-aligned rewards, MolmoAct MolmoAct (Lee et al., [43] reasons through spatial traces, Fast-ThinkAct Fast-ThinkAct (Huang et al., [32] compresses reasoning into latent thoughts, and ACoT-VLA ACoT-VLA (Zhong et al., [83] predicts coarse reference actions to guide policy generation. However, these methods largely augment robot demonstrations with intermediate reasoning, without explicitly grounding the learned thoughts in the task-relevant changes they induce in the world. Their generalization to novel scenes and skills therefore remains constrained by the behavioral coverage of the underlying robot data.

Learning Robot Policies from Human Videos. Improving generalization also requires expanding the behavioral coverage of the training data. Human videos capture substantially greater diversity in scenes, objects, and skills than robot demonstrations, motivating their use for scaling robot learning [74, 82, 44] (Kim et al. 2026). Supervised methods convert human interactions into robot-compatible representations, as in EgoScale EgoScale (Zheng et al., [82] and EgoWAM EgoWAM (Li et al., [44], but require costly cross-embodiment supervision. Latent action models instead infer action-like variables from visual transitions without explicit control labels [76, 59] (Chen et al. 2025), with LAPA LAPA (Ye et al., [76] learning discrete latent actions before grounding them into robot control. Such representations may capture appearance or embodiment cues rather than task-relevant dynamics. ThinkWorld grounds shared motion tokens and refines them by matching predicted and observed futures in videos.

World Models and Visual Foresight for Embodied Control. Recent advances in generative video models [70, 1] have enabled their application to a wide range of tasks [73, 68, 31, 38]. World models learn action-conditioned dynamics to support planning and policy optimization. Dreamer-style agents optimize policies through imagined rollouts [28, 29], while recent large-scale models acquire broad physical priors from diverse video data [2, 27] (Li et al. 2026). In robotics, visual-foresight methods generate video plans or image subgoals to guide control [24, 5, 81] (Black et al. 2024), while predictive VLAs forecast future pixels or latent features for planning and representation learning [80, 12] (Sun et al. 2026; Li et al. 2026; Kim et al. 2026). More recent world-action models jointly model future-world evolution and robot actions [77, 56, 13] (Kim et al. 2026). In contrast, ThinkWorld does not use future prediction as the control output or an imagined rollout for policy optimization. Instead, it treats agreement between predicted and observed futures as a training-time signal to reinforce action-grounded Chain-of-Dynamics within the VLM. This directly turns observed world changes, including those in action-unlabeled videos, into supervision for action reasoning, while future prediction is discarded at deployment.

Method

Method

Problem Formulation

Given a language instruction $l$, a current visual observation $o_t$, sampled history frames $h_t$, and proprioceptive state $s_t$ in case of robot, our goal is to predict a continuous robot action sequence $a_t$. Rather than treating actions solely as supervision for continuous control, ThinkWorld turns them into a reasoning interface for modeling how an action should unfold and transform the world. Through supervised fine-tuning, we instantiate this interface as a Chain-of-Dynamics (CoD) $c_t$, grounding its discrete motion vocabulary with action-labeled demonstrations across human and robot embodiments while learning CoD-conditioned future imagination through learnable queries that predict the future visual state (§ 4.2). We then reinforce CoD reasoning through world imagination, using the alignment between the imagined future and the observed future $o_{t+\Delta}$ after a physical horizon $\Delta$ as a learning signal that does not require motion annotations (§ 4.3). Together, action grounding and reinforcement through world imagination refine the model representations used for continuous control while extending learning to action-unlabeled videos at scale.

Chain-of-Dynamics for Embodied Action Reasoning

Robot actions are typically represented in embodiment-specific control spaces, such as joint positions or end-effector poses. Although directly executable, these control targets keep action supervision tied to individual embodiments, limiting transfer across robots and from human demonstrations. We therefore introduce Chain-of-Dynamics (CoD), a sequence of discrete motion tokens that provides a shared action language for human and robot motion, as illustrated in Figure 2(a).

Action-Grounded Reasoning Space. Let $m_t$ denote a motion segment from a robot trajectory or an annotated human hand trajectory. We map both to a normalized delta end-effector representation and tokenize them with a FAST-style motion tokenizer $\mathcal{T}(\cdot)$ (Pertsch et al., [57]), obtaining the target CoD sequence $c_t^* = \mathcal{T}(m_t)$ in a shared discrete vocabulary. Conditioned on the instruction, visual history, and current observation, the VLM learns to generate $c_t$ autoregressively. This tokenization provides a common interface grounded in motion from both sources without assuming kinematic equivalence between humans and robots.

CoD-Conditioned Future Imagination. Motion grounding defines a CoD sequence, but not the world change it induces. To predict this change, we append learnable future queries $q_{t+\Delta}$ after the CoD. Given the visual context and teacher-forced CoD $c_t^*$, the VLM contextualizes these queries as $\tilde{q}_{t+\Delta}$, which a lightweight decoder $\mathcal{D}_{\mathrm{future}}$ maps to the predicted future representation $z_{t+\Delta}$ at physical horizon $\Delta$:

Instead of reconstructing future RGB pixels, we predict in a frozen DINO feature space. DINO features retain dense spatial and object structure while abstracting away appearance details (Zhou et al., [84]; Siméoni et al. [63]), allowing imagined and observed futures to be compared by structural change without photorealistic generation. The target $z^*_{t+\Delta}$ is obtained by encoding the observed future frame $o_{t+\Delta}$ with the frozen DINO encoder $\mathcal{E}_{\mathrm{future}}$. This frame is used only to construct the target and is never provided as VLM input.

Joint Supervision of CoD and Future Prediction. We jointly train the VLM to generate CoDs from observed motion and predict future visual states conditioned on those CoDs. The objective is

where $\mathcal{L}_{\mathrm{CoD}} = \mathrm{CE}(p_\theta(c_t \mid l,h_t,o_t),c_t^*)$, $\mathcal{L}_{\mathrm{future}} = \operatorname{dist}(z_{t+\Delta},z^*_{t+\Delta})$.

Here, $\operatorname{dist}(\cdot,\cdot)$ is a feature distance metric and $\lambda_{\mathrm{future}}$ weights the future prediction loss. For continuous control, the action expert uses the VLM representations shaped by these objectives, rather than decoding or directly consuming CoD tokens (Intelligence et al., [34]; Chen et al. (2025); Sun et al. (2026)). Unlike CoD targets, future prediction targets are available from any temporally ordered human or robot video. In § 4.3, we use observed futures to evaluate sampled CoDs and extend learning to transitions without motion annotations.

Scaling Chain-of-Dynamics Reasoning via Reinforcing World Imagination

Once trained with joint supervision, the model generates CoDs and predicts future visual states conditioned on them. However, prediction under teacher-forced CoDs alone does not train the VLM to plan the intermediate dynamics needed to reach a desired future state. Therefore, we reward sampled CoDs according to the agreement between their imagined and observed futures, encouraging the VLM to reason about the world changes implied by its CoDs, as shown in Figure 2(b). Since this reward requires only observed transitions, it enables the VLM to explore alternative CoDs and learn from human and robot videos without requiring action labels or executing candidate motions.

Reward Shaping from World Imagination. For each candidate CoD, ThinkWorld predicts its future visual representation and compares it with the observed future in latent space. We compute the reward using a motion-enhanced feature similarity score $\operatorname{sim}(\cdot,\cdot)$, which places greater weight on DINO feature agreement in patches with larger current-to-future changes:

The score retains a uniform contribution from the full scene to preserve global consistency while preventing redundant background regions from dominating the reward (Shah et al., [61]). Motion weighting uses only observed frames, so scoring requires no action annotations.

Optimizing CoD Reasoning with World-Imagination RL. We aim to learn CoDs that capture the intermediate dynamics leading from the current scene to a desired future state. To train this reasoning process, ThinkWorld samples alternative CoDs from the same visual context and rewards those whose imagined futures best match the observed future. These rewards let ThinkWorld explore alternative CoDs that could lead from the current scene to the observed future, allowing passive human and robot videos to guide action reasoning without action labels or online execution of candidate motions.

We update the VLM with GRPO (Shao et al. [62]) using normalized world-imagination rewards (Liu et al. (2026)), while supervised replay preserves the motion grounding learned during SFT:

By comparing imagined and observed futures, this stage encourages the VLM to encode action-relevant dynamics in the representations later used for continuous control. All other components remain frozen during RL, and neither CoD generation nor future prediction is required at deployment. Complete optimization details are provided in Appendix § 7.6.

Experiment

Implementation Details

Training Setup. For the motion tokenizer $\mathcal{T}(\cdot)$, we convert both robot actions and human wrist poses into a unified delta end-effector representation. We use Qwen3-VL-2B (Bai et al. [4]) as the VLM backbone with the action expert from MolmoAct2 (Fang et al. [25]). Future queries are decoded by a Transformer decoder, with ground-truth future representations extracted using DINOv3 (Siméoni et al. [63]). Our training proceeds in two stages. During SFT and RL pre-training, we first fully fine-tune the VLM for CoD reasoning and future prediction, and then adapt the action expert while freezing the VLM following Driess et al. [23]. We subsequently perform RL with $K = 4$ sampled trajectories, updating only the VLM while keeping the action expert frozen. We use 128 and 64 NVIDIA H100 GPUs for SFT and RL, respectively. During Embodiment-Specific Post-Training, we further adapt the pretrained model to each target embodiment. ThinkWorld-LIBERO is trained on the standard LIBERO training set to evaluate generalization and robustness to novel environments and perturbations, while ThinkWorld-YAM is adapted on MolmoAct2-BimanualYAM for real-robot experiments. We use 32 and 64 NVIDIA H100 GPUs for the two settings, respectively.

Training Datasets. Our training corpus combines human egocentric data from EgoDex (Hoque et al. [30]) and EgoVerse (Punamiy­a et al. [58]) with robot demonstrations from MolmoAct2-DROID/BimanualYAM (Fang et al. [25]), Bridge/RT-1 (O’Neill et al. [55]), BC-Z (Jang et al. [35]), and RoboCOIN (Wu et al. [72]). During SFT, we also use auxiliary VQA data to preserve visual-language capabilities and reduce catastrophic forgetting. Wrist views are included when available; otherwise, only the main view is used.

Real-Robot Experimental Setup. We use a bimanual YAM system equipped with three cameras for real-robot evaluation: one global top-view camera and wrist-view camera for each arm. We use only RGB observations for both training and inference. The embodiment-adapted ThinkWorld-YAM model is then used as the initialization for efficient task-specific fine-tuning. Additional implementation and training details are provided in Appendix.

Real-World Generalization under Distribution Shifts

We evaluate real-world generalization under matched in-domain (ID) and out-of-distribution (OOD) settings across three manipulation tasks: placing cups on a hanger, color matching, and object sorting. OOD settings introduce changes in object color or category and, in some cases, require trajectories not observed during task-specific training. For each task, we collect demonstrations across at least six object configurations and evaluate multiple settings under both ID and OOD conditions, with five trials per setting. These tasks require the policy to identify task-relevant object attributes, such as color and category, and adapt its motion accordingly.

As shown in Figure 3, we compare ThinkWorld-YAM with Diffusion Policy [19], MolmoAct2 [25], a VLA with explicit action reasoning, and LaWAM [15], which conditions action generation on predicted latent visual subgoals. While the methods perform similarly in-domain, their differences become more pronounced under OOD shifts. ThinkWorld maintains higher success rates when object appearances change or unseen trajectories are required, outperforming both MolmoAct2 and LaWAM. This suggests that grounding CoD reasoning in action-induced world changes helps the policy adapt to task-relevant changes rather than relying on appearance cues or memorized trajectories.

Acquiring Skill Composition from Human Experience

We further evaluate whether human demonstrations can transfer compositional knowledge to robot control. We construct an oven manipulation task from reusable atomic robot skills and provide robot demonstrations only for the individual skills 1) open, 2) close, 3) put the left object in the oven, and 4) put the right object in the oven, while a human video demonstrates how a complete compositional task can be achieved. Existing end-to-end VLA approaches typically learn executable control from robot action supervision [25], making it difficult to compose complete robot tasks from human demonstrations when only atomic robot skills are supervised. Successful execution requires the robot to identify the relevant skills, preserve their temporal order, and transition between them without observing the complete robot trajectory during training. As shown in Figure 4, a human demonstration of open $\to$ put-left $\to$ close enables ThinkWorld to compose the corresponding robot skills and execute the full sequence in a single rollout. These results demonstrate that the shared CoD reasoning space enables compositional knowledge to transfer from human demonstrations to robot execution, allowing the robot to perform a complete skill sequence never demonstrated in robot data. Additional qualitative examples of skill composition in drawer manipulation are provided in the Appendix § 9.§ 1.

Robustness to Task and Environment Shifts in Simulation

We evaluate ThinkWorld on standard LIBERO [49] and the more challenging LIBERO-PRO [85]. LIBERO evaluates manipulation across spatial, object, goal, and long-horizon task suites, while LIBERO-PRO introduces controlled shifts in object attributes, positions, instruction semantics, task goals, and environments. ThinkWorld-LIBERO is post-trained only on the standard LIBERO training set and evaluated directly on LIBERO-PRO without exposure to these perturbations. As shown in Table 1, ThinkWorld achieves the best overall performance on both benchmarks, with average success rates of 97.3% on LIBERO and 55.6% on LIBERO-PRO. Notably, under task and environment shifts, ThinkWorld reaches 31.9% and 60.9%, outperforming VLA-JEPA by 11.1 and 19.9 percentage points, respectively. While VLA-JEPA also incorporates latent world modeling, these gains suggest that grounding CoD reasoning in action-induced world changes provides more transferable structure for adapting to shifts beyond the training distribution.

Model	Params.	Spatial	Object	Goal	Long	Avg.	Object	Position	Semantic	Task	Environment	Avg.
OpenVLA (Kim et al., [41])	7.5	84.7	88.4	79.2	53.7	76.5	<u>93.0</u>	0.0	97.3	0.0	68.0	51.7
$\pi_0$ (Black et al., [7])	3.2	96.8	98.8	95.8	85.2	94.2	90.5	0.0	90.5	0.0	38.8	44.0
$\pi_{0.5}$ (Intelligence et al., [34])	3.4	98.8	98.2	98.0	92.4	96.9	96.0	20.8	<u>95.8</u>	0.8	52.8	<u>53.2</u>
MolmoAct (Lee et al., [43])	8.1	87.0	95.4	87.6	77.2	86.6	76.0	1.5	85.8	1.5	–	41.2
VLA-JEPA (Sun et al., [64])	2.8	96.2	99.6	97.2	95.8	<u>97.2</u>	80.4	3.5	97.2	20.8	41.0	48.5{sup}†`
ThinkWorld (Ours)	2.9	<u>98.2</u>	<u>98.8</u>	<u>97.8</u>	<u>94.2</u>	97.3	73.3	<u>18.2</u>	93.9	31.9	<u>60.9</u>	55.6

Analysis

Cross-Embodiment Structure of CoD. We first examine whether CoD induces a shared motion-centric structure across human and robot embodiments. For each sample, we extract the VLM hidden state of the final </think> token, which summarizes the generated CoD, and visualize these features with t-SNE in Figure 5. We use bimanual human and robot datasets, with colors denoting four 2D end-effector motion directions and marker shapes indicating dataset sources. Despite substantial differences in embodiment and visual appearance, samples with similar motion directions form coherent clusters that contain examples from multiple datasets, while samples from the same dataset separate according to their motions. This suggests that CoD features are organized primarily around action-relevant motion rather than embodiment identity, revealing a shared action-level structure shared across human and robot behaviors.

Effectiveness of Human Data. We investigate the contribution of human egocentric video through a controlled pretraining study, keeping the amount of robot data fixed while increasing the human video fraction to 10%, 25%, and 50% of the total training data. We compare these variants against a robot-only baseline and an additional robot-only control that matches the training scale and number of training iterations of the 50% human-data variant by using additional robot data. This control helps distinguish the benefits of human video from those of increased training scale alone. These experimental configurations are designed to assess the contribution of heterogeneous human video to robustness and differ from our final model configuration. After SFT, each variant undergoes embodiment-specific post-training on LIBERO and is evaluated on LIBERO-PRO to assess generalization. As shown in Figure 6, overall success increases monotonically with the proportion of human video, from 45.4% for the robot-only baseline to 47.8% for the 50% human-data variant. The latter also outperforms the matched robot-only control by approximately 1.2 percentage points. These results suggest that the diversity provided by human egocentric video improves generalization beyond the gains attributable to increased training scale alone.

Role of CoD in Future Prediction. We further examine how CoD contributes to future prediction. As shown in Figure 7, predictions conditioned on the generated CoD better capture the action-induced transition toward the ground-truth future, whereas shuffled, empty, or no-CoD variants tend to preserve the current-state structure or predict incorrect transitions. Table 2 quantitatively supports this observation: CoD achieves a cosine similarity of 0.74 across all patches, compared with 0.71–0.72 for the other conditions. The difference is more pronounced in the top 25% most dynamic target patches, where CoD reaches 0.62 compared with 0.54–0.55, indicating that CoD is particularly effective at predicting action-induced changes.

Condition	All patches		Top-25% motion
	$\mathrm{Cos}\uparrow$	$\Delta$	$\mathrm{Cos}\uparrow$	$\Delta$
Shuffle	0.71	−0.024	0.55	−0.071
Empty	0.72	−0.020	0.54	−0.079
CoD off	0.72	−0.018	0.55	−0.071
CoD on	0.74	—	0.62	—

We further examine what information CoD captures through attention visualization in Figure 8. Across both human and robot examples, CoD consistently attends to action-relevant dynamic regions, including human hands or robot arms and the manipulated objects, rather than static background regions. This pattern indicates that CoD focuses on the parts of the scene most involved in the ongoing motion and links these motion cues to the resulting changes in object state and configuration. Together with the future-prediction results, these visualizations suggest that CoD captures action-centric dynamics that ground future prediction in the state transitions induced by an action.

Effect of Training Components. To assess the effectiveness of each training component, we start from a baseline consisting of the Qwen3-VL backbone with the MolmoAct2 action expert, and progressively add visual history, future prediction, CoD, and world-imagination RL. As shown in Table 3, history provides temporal context and slightly improves the average success rate from 40.1% to 40.7%. Future prediction brings a much larger gain to 50.4%, indicating the benefit of explicitly modeling how the scene evolves. Adding CoD further raises performance to 54.4%, with consistent improvements across all distribution shifts, showing the value of grounding future prediction in action-relevant dynamics. Finally, world-imagination RL improves the average to 55.6%, with further gains under position, task, and environment shifts, demonstrating its role in refining CoD reasoning beyond supervised training. Overall, these ablations support the design of ThinkWorld, showing that each component contributes to more robust and generalizable action reasoning.

Model variant	Object	Pos.	Sem.	Task	Env.	Avg.
Qwen baseline	59.5	1.9	84.4	17.2	37.8	40.1
+ History frames	61.3	4.9	84.6	22.8	29.8	40.7
+ Future prediction	70.0	11.9	89.4	26.3	54.4	50.4
+ CoD	74.5	16.9	94.4	29.7	56.4	54.4
+ RL	73.3	18.2	93.9	31.9	60.9	55.6

Conclusion

We present ThinkWorld, a VLA learning framework that learns transferable Chain-of-Dynamics (CoD) through reinforcement learning with world imagination. By grounding action thoughts in the world changes they induce, ThinkWorld learns a common representation across human and robot embodiments and extends action reasoning to videos without explicit action annotations. Extensive experiments demonstrate effective human-to-robot transfer and strong generalization to unseen tasks and environments, validating the effectiveness of learning transferable action reasoning from observed world dynamics and its potential for scaling robot learning with diverse human and robot experience.

Extended Implementation Details

Model Configuration

ThinkWorld uses Qwen3-VL-2B-Instruct [4] as its vision-language backbone and a frozen DINOv3 ViT-H/16+ target encoder [63]. The DINO encoder receives the main-camera target image at $224 \times224$ resolution and produces a $14 \times14$ grid of 1,280-dimensional patch features. We append 64 learned future-query tokens to the VLM sequence. A two-layer MLP and a six-layer cross-attention decoder map their 2,048-dimensional hidden states to 196 predicted DINO patch features. The future target is the main-camera observation closest to exactly one second after the current observation; samples more than 50 ms from the requested time are masked from the future loss.

Continuous actions are produced by an action expert initialized from MolmoAct2 [25]. The expert contains 36 non-causal Transformer blocks with hidden width 768 and operates in FP32. It exposes a maximum $30 \times32$ action tensor, but a two-dimensional validity mask restricts each example to its dataset-native horizon and action channels. The expert receives layer-wise Qwen key/value features from input tokens only; generated CoD tokens and future-query tokens are not part of its direct context.

For clean normalized action $a$ and Gaussian noise $\epsilon$, we sample $z \sim\operatorname{Beta}(1, 1.5)$, set $t = 0.001 + 0.999z$, and construct $x_t = (1 - t)\epsilon+ ta$. The action expert regresses $a - \epsilon$ using masked mean-squared error over valid entries. We draw eight independently noised flow samples per training example and use ten Euler integration steps at inference.

Input and Output Configuration

Input prompt. All media precede the textual metadata in a single user turn. When temporal context is available, the order is main-view history, current main view, and the remaining current views in their dataset-defined order. History contains at most 16 main-view frames sampled at 2 Hz and ordered from oldest to newest. The current main image is not repeated in the history.

The complete VLA input/output template is shown below. Bracketed lines are optional, and the textual portion is rendered as one normalized string rather than as separate lines.

Future-prediction prompt template

USER MEDIA:
[main-camera history video: 0--16 frames]
<current_main_image>
[<current_wrist_image_1><current_wrist_image_2>]

USER TEXT:
The task is to {task}. The setup is {embodiment}
[ with {native_control} control].
[The current state of the robot is {state}.]
[A {N}-frame history video is followed by the current observation image.
| The current observation frame is provided.]
The future visual target is the main-view observation exactly one second
after the current observation. Given these, produce a future motion plan
and then predict the future state one second in the future.

ASSISTANT:

Proprioceptive state encoding. Following MolmoAct2 [25], we encode proprioceptive state as discrete tokens. Each dataset-native state dimension is normalized using training-split statistics, clipped to $[-1, 1]$, uniformly quantized into 256 bins, and serialized between dedicated state-boundary tokens. The native state width is preserved, so no common padded state vector is exposed in the prompt.

Media processing. Dataset-side resizing is disabled. Qwen retains the source aspect ratio, with a pixel-area range of 4,096–200,704 for each current image and 4,096–50,176 for each history frame. State conditioning is dropped with probability 0.10 and wrist views with probability 0.20 during supervised training. The maximum multimodal sequence length is 2,048 tokens; over-length examples are removed rather than silently truncating their supervision.

Training-time image augmentation can be summarized with the following torchvision-style transform:

Image augmentation pipeline

import torchvision.transforms as T

transform = T.Compose([
    T.RandomResizedCrop((height, width), scale=(0.95, 0.95)),
    T.RandomRotation(degrees=5),
    T.ColorJitter(brightness=0.2, contrast=0.2,
                  saturation=0.2, hue=0.05),
    T.RandomApply([
        T.GaussianBlur(kernel_size=5, sigma=(0.1, 1.0))
    ], p=0.2),
])

The transform parameters are sampled once and replayed across all views and time steps in an example. RL disables image augmentation so that the four sampled completions for a prompt receive identical media.

CoD tokens receive language cross-entropy supervision, whereas future-query tokens are excluded from cross-entropy and supervise only the future projector. Action SFT uses an empty <think>...</think> wrapper and disables assistant-token cross-entropy.

Motion Tokenizer Training. We train a motion tokenizer following the FAST architecture [57] on one-second windows of consecutive wrist-motion deltas from robot and human trajectories. Each time step contains a 6D motion vector for one hand or a 12D vector for two hands: translation in meters and rotation as a rotation vector in radians, with gripper state excluded. We normalize translation and rotation using fixed scales of 0.3448 m and 0.9119 rad, respectively, and encode the sequences with a vocabulary of 2,048 tokens. Robot poses are expressed in the robot base frame, whereas human poses are expressed in an actor-centered frame whose $+z$ axis is gravity-aligned and whose $+x$ axis follows the actor’s current yaw. Training on both domains yields a shared token space for intended wrist motion. The tokens are not executed as native robot actions; a separate continuous action expert predicts executable actions.

Training Data

VLM SFT combines robot trajectories from BimanualYAM [25], DROID [37], RoboCOIN [72], Bridge [69], RT-1 [55], and BC-Z [35] with egocentric human-motion trajectories from EgoVerse [58] and EgoDex [30]. Table 4 summarizes the frozen source scale and native action contracts without exposing dataset sampling weights. Episode and frame counts describe source data before the episode-disjoint validation holdout, and a frame denotes one trajectory timestep rather than one image per camera. For Action SFT, we exclude the 50-Hz RoboCOIN trajectories and retain only its 30-Hz trajectories.

Source	Episodes	Frames	Hours	Hz	Camera view	Action representation
YAM	32.2K	76.0M	704.1	30	Top + two wrists	$30 \times14$D absolute joint
DROID	74.6K	17.8M	328.9	15	Exterior + wrist	$15 \times8$D absolute joint
RoboCOIN	112.0K	82.3M	755.2	30/50	One or three views	30/50 $\times$ (12/14)D absolute EEF
Bridge	53.2K	1.9M	105.2	5	Main	$5 \times7$D delta EEF
RT-1	87.2K	3.8M	350.6	3	Main	$3 \times7$D delta EEF
BC-Z	39.4K	5.5M	152.0	10	Main	$10 \times7$D delta EEF
EgoVerse	55.6K	124.8M	1,155.4	30	Egocentric	One-second 6D/12D motion tokens
EgoDex	314.8K	78.3M	724.8	30	Egocentric	One-second 6D/12D motion tokens
Total	769.1K	390.3M	4,276.1	–	Mixed	Dataset-native

Auxiliary visual-language supervision covers image and video captioning, caption question answering, spatial reasoning, pointing, counting, point tracking, robotics-oriented pointing, and robot VQA. Specifically, Pixmo-Cap [22], Molmo2-ER [25], Molmo2 Cap-LLaVA, Molmo2 Video-Track [20], RoboPoint [78], and RoboVQA [60]. These examples retain their language targets and do not contribute continuous-action or future-feature losses.

Training Configuration

Training consists of VLM SFT, Action SFT, and RL. All phases use fused AdamW with $(\beta_1,\beta_2) = (0.9, 0.95)$, $\epsilon= 10^{-8}$, weight decay 0.1, gradient-norm clipping at 1.0, and DeepSpeed ZeRO stage 1 without optimizer offload. Table 5 shows the detailed configuration by phase for direct comparison.

Setting	VLM SFT	Action SFT	RL
Initialization	Qwen3-VL-2B	VLM SFT	Action SFT
Trainable modules	Language model + future projector	Action expert	Language model
Objective	CoD/VQA CE + $3\mathcal{L}_{\mathrm{future}}$	Masked flow matching	GRPO/GDPO + weak SFT replay
Data	Robot + human motion; auxiliary VQA	Robot trajectories	Robot + human motion
Iterations	150k	50k (5k current-view + 45k history)	2k
Learning rate	VLM $10^{-5}$; projector $10^{-4}$	Expert $10^{-4} \to5 \times10^{-5}$	$10^{-6}$
Schedule	5% warmup, constant	Warmup + cosine cooldown	50-step warmup, constant
Batch	128	128	128
Seed	42	43 / 44	45
Allocation	16 × 8 GPUs	8 × 8 then 16 × 8 GPUs	8 × 8 GPUs

VLM SFT and RL execute the VLM in BF16. During Action SFT, Qwen uses BF16 autocast while the action expert is stored and optimized in FP32. For downstream LIBERO evaluation, we full-finetune the final model on demonstrations from the Object, Goal, Spatial, and LIBERO-10 suites for 50k iterations using masked flow matching, cosine learning-rate decay.

Details on Future Prediction Loss Design

Details on Future Prediction Loss Design

Following the main text, let $z_{t+\Delta}$ be the predicted DINO feature map and $z^{*}_{t+\Delta} = \mathcal{E}_{\mathrm{future}}(o_{t+\Delta})$ the target from the observed future, with $\Delta= 1$ s. The distance used in $\mathcal{L}_{\mathrm{future}} = \operatorname{dist}(z_{t+\Delta}, z^{*}_{t+\Delta})$ is

where $\mathcal{V}$ contains the valid image patches. Both feature maps are $\ell_2$-normalized before computing the loss. We use $\beta= 0.01$ and $\lambda_{\mathrm{SL1}} = 0.25 \times21.792413$. During RL, the future projector is frozen and $\mathcal{L}_{\mathrm{future}}$ is not optimized.

Details on RL Training Objectives

For each context, we sample $K = 4$ CoDs and use the frozen future projector to obtain a predicted future representation from each candidate. The similarity function in the main-text reward $r_{\mathrm{world}} = \operatorname{sim}(z_{t+\Delta}, z^{*}_{t+\Delta})$ averages global and motion-weighted patch agreement:

where $s_p = \cos(z_p,z_p^*)$. To form $w_p$, we measure the DINO feature change between the current and observed future frames. The weight is 0.2 for patches at or below the median change, increases linearly to 1 at the 90th percentile, and is clipped outside that range. For nearly static targets, we use the global average for both terms.

A malformed CoD receives a reward 0.05 below the lowest candidate similarity in its group. We standardize rewards within each four-candidate group and then across the rollout batch; each candidate’s resulting advantage is shared by its tokens. GRPO uses token-level importance ratios clipped to [0.8, 1.28] and a KL coefficient of 0.04. In the main-text objective $\mathcal{L}_{\mathrm{RL}}$, we set $\lambda_{\mathrm{GRPO}} = 0.999$ and $\lambda_{\mathrm{replay}} = 0.001$. The replay term is teacher-forced cross-entropy on a separate minibatch of canonical CoDs. Only the VLM is updated during RL; the action expert and future projector remain frozen.

Evaluation Details

Simulation Benchmarks and Baselines

We fine-tune our policy only on standard LIBERO (Liu et al., [49]) and use its unmodified tasks as the in-distribution reference. We then evaluate the trained policy on LIBERO-PRO and LIBERO-Plus without further adaptation to either benchmark. We select LIBERO-PRO as the primary benchmark because its controlled changes to objects, object positions, task specifications, and environments directly test whether the policy can adapt its behavior to new manipulation settings. LIBERO-Plus additionally covers camera, lighting, background, and sensor-noise perturbations, which probe perceptual robustness beyond our main methodological focus. We report a supplementary LIBERO-Plus evaluation in Appendix § 9.§ 5.

Real-Robot Setup

We use the BimanualYAM robot system to collect our in-lab dataset and evaluate trained policies. As shown in Figure 9, the system comprises two YAM arms and three RealSense D405 cameras: one providing a top view and one mounted on each wrist. Policies use RGB observations only.

Policy inference runs on a single NVIDIA RTX 3090. To account for differences in inference latency, we query each policy and execute its actions at the same 15 Hz control rate, while the remaining system components run at 30 Hz. For each predicted action chunk, we execute the first half and then replan using the latest observation.

Real-Robot Evaluation

For the real-robot experiments, we evaluate three tasks on the bimanual YAM platform: color matching (60 robot demonstrations), cup hanging (45), and object sorting (100). Within each task, the demonstrations span multiple trajectories and environment configurations, and we use them jointly to fine-tune a single task-specific policy. The tasks are designed to test whether the policy can infer task-level regularities and adapt its actions to the current scene rather than memorize a single trajectory. Separately, we study skill composition on the oven task using 100 robot episodes of atomic-action

We fine-tune ThinkWorld for 50,000 steps, applying LoRA to the VLM and fully updating the action DiT decoder. ThinkWorld can also leverage human videos during training; for these samples, we optimize only the VLM future-prediction loss and omit the action DiT loss. We train each baseline for 50,000 steps using its model-specific fine-tuning recipe from the official LeRobot (Cadene et al., [11]) repository.

Additional Analyses and Results

Detailed Real-Robot Evaluation Results

In-domain and out-of-domain task evaluation. Table 6 reports detailed rollout success rates under in-distribution (ID) and out-of-distribution (OOD) conditions for all tasks. ThinkWorld achieves the highest success rate in all six task–condition pairs, succeeding in all 60 ID trials and 52 of 60 OOD trials. Its macro and micro success rates are 94.4% and 93.3%, respectively, compared with 62.2% and 60.0% for the strongest baseline, MolmoAct2. Example in-domain and out-of-domain rollout trajectories for all compared methods are shown in Figs. 10 (b) and (c) respectively.

Model	Color Matching ID	Color Matching OOD	Cup Hanging ID	Cup Hanging OOD	Sorting ID	Sorting OOD	Macro SR	Micro SR
Diffusion Policy	18/30 (60.0)	11/30 (36.7)	9/15 (60.0)	1/15 (6.7)	4/15 (26.7)	1/15 (6.7)	32.8	36.7
MolmoAct2	24/30 (80.0)	8/30 (26.7)	11/15 (73.3)	14/15 (93.3)	12/15 (80.0)	3/15 (20.0)	62.2	60.0
LaWAM	16/30 (53.3)	8/30 (26.7)	9/15 (60.0)	12/15 (80.0)	4/15 (26.7)	4/15 (26.7)	45.6	44.2
ThinkWorld (ours)	30/30 (100.0)	24/30 (80.0)	15/15 (100.0)	15/15 (100.0)	15/15 (100.0)	13/15 (86.7)	94.4	93.3

Skill composition. Fig. 11 presents qualitative results for real-robot skill composition in oven and drawer manipulation tasks. Each task involves opening the oven or a specified drawer, optionally placing the left or right object inside, and closing it. Robot demonstrations cover only the individual atomic skills (OPEN, PUT-L, PUT-R, and CLOSE; panel (a)), while human demonstrations specify complete skill sequences (panel (b)). The oven tasks include open $\to$ close and open $\to$ put-left $\to$ close. The drawer tasks include open $\to$ put-left $\to$ close and open $\to$ put-right $\to$ close for the upper drawer, and open $\to$ put-left $\to$ close for the lower drawer. The successful rollouts in panel (c) show ThinkWorld executing these compositions under the corresponding natural-language instructions, manipulating the specified objects and compartments while preserving the required skill order. These qualitative results illustrate the transfer of compositional knowledge from human demonstrations to robot execution without complete robot task trajectories during training.

Additional Qualitative Results on PCA and Attention Analysis

We provide additional qualitative visualizations that extend the analysis in the main text. Future-representation PCA and attention visualizations leverage Qwen layers 24–27. PCA and attention are descriptive analyses and are not used as causal evidence.

Fig. 12 expands the PCA visualization of future-prediction representations across five CoD conditions: the original CoD, shuffled CoD, empty CoD, a donor CoD from another episode, and no CoD. This comparison illustrates how the predicted future representations vary with CoD conditioning.

Figure 13 extends the CoD-to-input-observation attention maps shown in the main text and additionally shows attention from future-prediction tokens to the input observation. CoD attention tends to concentrate on dynamic, task-relevant regions, whereas future-prediction attention is distributed more broadly across the scene. This pattern suggests that CoD emphasizes localized changes relevant to action, while future prediction incorporates wider scene context.

Closed-Loop Future-Prediction Results

Table 7 reports SimplerEnv Li et al., [47] success rates without additional embodiment-specific fine-tuning for this benchmark. The model nevertheless encountered Bridge and RT-1 data during SFT and RL, providing prior experience with related robot settings. It achieves 49.31% on Google visual matching, 46.54% on the Google-VA variants (micro average), and 41.67% on WidowX visual matching, with an overall micro success rate of 46.51%. These results show that the policy can complete a substantial fraction of SimplerEnv rollouts without benchmark-specific adaptation.

Group	Task	SR (%)
Google-VM		49.31
	Pick Coke can	56.00
	Drawer	33.33
Google-VA	Move-near	55.83
	Put-in-drawer	2.12
	Micro avg.	46.54
WidowX VM		41.67
Overall	micro	46.51

Figure 14 visualizes future predictions during closed-loop SimplerEnv rollouts. In each column, the one-second-ahead prediction is paired with the observation subsequently reached at the predicted time. The predictions are PCA-projected feature maps and their spatial structure and progression closely track the corresponding observations across the three examples. This qualitative alignment suggests that the future representation remains informative as the policy encounters its own observations and replans under distribution shift.

LIBERO-Plus Evaluation Results

Table 8 reports zero-shot LIBERO-Plus ([26]) results using the policy fine-tuned only on standard LIBERO. ThinkWorld achieves the highest reported Avg. among the four methods at 71.7%, exceeding DreamVLA by 1.8 percentage points. It leads under robot initial-state, lighting, and object-layout perturbations. In particular, ThinkWorld reaches 70.5% on robot initial-state changes, compared with 40.8% for the next-best method, and 79.6% on object-layout changes, compared with 74.2%. These results are consistent with our goal of adapting actions to changed physical configurations and provide supplementary evidence of generalization beyond standard LIBERO.

Method	Camera	Robot	Language	Light	Background	Noise	Layout	Avg.
OpenVLA-OFT ([41])	56.4	31.9	79.5	88.7	93.3	75.8	74.2	69.6
$\pi$0-FAST ([7])	65.1	21.6	61.0	73.2	73.2	74.4	68.8	61.6
DreamVLA ([80])	65.0	40.8	63.5	85.7	82.6	84.9	74.0	69.9
ThinkWorld (Ours)	46.9	70.5	73.5	90.7	90.0	62.7	79.6	71.7

Inference Efficiency

Table 9 reports steady-state batch-1 policy latency with CUDA graph capture disabled, using eager inference without static-shape graph setup. All models are measured on the same NVIDIA H100 80GB GPU, with five warm-up calls followed by 50 CUDA-synchronized timed calls. Each model uses its native deployed input contract, so the table reflects practical policy latency rather than an identical-input architecture ablation.

Policy	p50 (ms)	p95 (ms)	Wall (s)	Mem. (GiB)
MolmoAct2	335.07	467.96	18.44	12.11
MolmoAct2-Think	879.37	1014.47	46.76	12.12
ThinkWorld	324.84	333.09	16.31	7.07

ThinkWorld uses the reasoning process only during training, where it encourages the VLM to learn more action-centric representations. The deployed policy does not execute this process at inference, so each policy call avoids its additional runtime computation. This separation offers an explanation for how training-time reasoning can support strong task performance while preserving low inference latency.

Limitations and Reproducibility

Training results are obtained from a single training seed; accordingly, the reported episode-level confidence intervals characterize evaluation variability rather than variation across independently trained checkpoints. For real-robot experiments, we report the number of physical trials and task-level results to make the evaluation scope explicit. Inference latency should be interpreted under the stated deployment setup, as it varies with hardware and input configuration. To support reproducibility, we provide trial denominators, model initialization details, training configurations, and hardware specifications alongside the aggregate results.