WorksComputer Science
Wireless Edge-Assisted Adaptive Latent Reasoning for Vision-Language-Action Models in Physical AI
Abstract
Vision-language-action (VLA) models enable embodied artificial intelligence (AI) by translating visual observations and language instructions into robotic actions. Emerging use cases in physical AI are expected to drive the need for efficient inference and communication, as the computational demands of the VLA models pose challenges for real-time control with limited onboard resources. Test-time scaling through latent reasoning allows these models to refine hidden representations without generating explicit chain-of-thought (CoT) tokens, but introduces a trade-off between task success, inference latency, and energy consumption. In this paper, we address the efficient collaborative inference under a realistic communication scenario. We propose a wireless edge-assisted framework for adaptive latent reasoning (WEALR) that augments onboard computation with edge computing. A decision-making agent jointly selects the execution location and latent-reasoning convergence threshold. We formulate this joint adaptation as a constrained Markov decision process (CMDP). A Lagrangian-constrained deep reinforcement learning (CDRL) approach, namely trust region policy optimization Lagrangian (TRPO-Lag), learns to balance task performance and resource consumption under varying wireless conditions and task requirements. We evaluate the framework using LIBERO manipulation tasks, profiled VLA inference costs, and ray-traced wireless channels. Experimental results demonstrate that TRPO-Lag achieves up to 23% higher feasibility-weighted task success rate compared to benchmark CDRL methods.
Introduction
Physical artificial intelligence (AI) brings perception, reasoning and action into embodied systems that interact with the physical world [1]. Vision-language-action (VLA) models, such as OpenVLA [2], support this capability by translating visual observations and language instructions into robotic actions. Advances in these models create opportunities for generalizable physical AI, but their computational demands challenge deployment on resource-constrained robots. Task performance must therefore be balanced against onboard computation, energy consumption, and latency requirements [3], [4]. In particular, timely robotic control requires communication and inference to jointly satisfy a latency service-level objective (SLO). We refer to robots capable of visual perception, instruction understanding, planning, reasoning, and action execution as physical AI devices (PADs) in this paper.
Edge computing augments limited onboard computation by enabling PADs to offload inference to nearby edge servers, but its benefits depend on wireless propagation conditions
[5]. This deployment perspective aligns with the vision of using radio access network (RAN) infrastructure to support AI applications at the network edge, such as mobile switching offices (MSO) [6]. However, faster edge computation comes at the cost of wireless transmission delay. In indoor environments such as homes and warehouses, blockage and multipath propagation create heterogeneous line-of-sight (LOS) and non-line-of-sight (NLOS) conditions. Changes in PAD position affect achievable uplink (UL) rates and can offset the computational advantage of edge execution. Effective offloading must therefore account for wireless conditions alongside computation and control-latency requirements. Evaluating this tradeoff requires environment-specific channel models that capture blockage, multipath propagation, and spatial variations in link quality, which ray tracing can provide [7], [8].
Reliable physical interaction requires VLA models to interpret diverse visual observations and language instructions and generate appropriate actions across tasks of varying difficulty. Supporting these capabilities motivates increasing model capacity through parameter scaling or allocating additional computation during inference. Sparse mixture-of-experts (MoE) architectures enable parameter scaling while limiting computation per input by activating only a subset of model parameters [9]. However, their memory and computational requirements can still challenge deployment on resource-constrained PADs. Test-time scaling offers a complementary approach that adjusts computational effort within a fixed model, making it particularly attractive for wireless edge deployment, where available resources and latency requirements limit inference costs [5].
Unlike explicit reasoning methods such as chain-of-thought (CoT), which generate intermediate text through sequential decoding, latent reasoning refines hidden representations without producing intermediate text tokens, potentially reducing the decoding overhead and latency associated with textual reasoning. The recurrent-depth architecture in [10] enables large language models (LLMs), which serve as backbones of VLAs, to vary computational depth through shared parameters, offering adjustable inference effort within a fixed model footprint. For robotic inference, recurrent-depth VLA (RD-VLA) applies a related principle through a weight-tied recurrent action head and a convergence-based stopping criterion to adapt the number of reasoning iterations across observations [11]. For example, grasping a small object may require more reasoning than moving through an unobstructed region. A single convergence threshold can therefore yield different reasoning depths across stages of a task and different tasks. Additional reasoning may improve action quality but also increases inference latency and resource consumption, motivating adaptive control of reasoning effort. These considerations motivate jointly selecting VLA execution location and reasoning effort under wireless, energy, and computation constraints.
Related studies address complementary aspects of collaborative inference. The device-cloud framework in [12] uses resource-constrained reinforcement learning (RL) to select LLM inference locations and multimodal inputs. Also, the EcoVLA approach proposed in [13] selects energy-efficient VLA co-inference schemes under real-time constraints, adapting to network fluctuations and edge workload variations. The authors of [14] propose MORES, a reasoning-as-a-service framework that distributes recurrent hidden-state updates between devices and servers, jointly adapting reasoning computation and latent-feature pruning under heterogeneous wireless conditions. While conversational LLM services typically prioritize response quality, VLAs must deliver actions within deadlines imposed by ongoing physical control [3].
Meeting these deadlines requires balancing task success against communication delay and resource consumption. Cumulative budgets couple decisions across requests, and shared edge resources couple decisions across PADs, motivating a constrained Markov decision process (CMDP) formulation. Constrained deep reinforcement learning (CDRL) addresses such problems by separating performance objectives from explicit operating limits [15]. This motivates the use of CDRL methods, including Lagrangian variants of proximal policy optimization (PPO) and trust region policy optimization (TRPO), as well as constrained policy optimization (CPO), for resource management in wireless networks. In particular, PPO-Lag has been adopted for joint unmanned aerial vehicle (UAV) altitude control and Internet of Things (IoT) channel access [16].
In this paper, we propose Wireless Edge-Assisted Adaptive Latent Reasoning (WEALR), a framework that jointly adapts VLA execution location and reasoning effort. For each robotic task request, an agent selects local or edge execution and a reasoning convergence threshold. This threshold governs termination of the latent reasoning loop while allowing the realized reasoning depth to vary across observations within the task. We use TRPO-Lag, which learns the policy and constraint multipliers, connecting adaptive latent reasoning with wireless offloading and cumulative resource management. The key contributions are as follows:
We develop WEALR to jointly select VLA execution location and reasoning threshold according to task requirements, wireless conditions, and available resources.
We formulate a CMDP with per-PAD energy, shared edge-reasoning, and SLO constraints, and apply TRPO-Lag to optimize robotic task success.
We evaluate task success, feasibility, and policy behavior using LIBERO manipulation tasks [17], profiled VLA inference costs, and ray-traced channels under varying wireless conditions and resource budgets.
System Model and Problem Formulation
We consider an edge-computing system comprising PADs and an edge server each equipped with AI acceleration hardware co-located with an indoor base station (BS), as shown in Figure 1 (a). Both PADs and the server host pretrained VLA models with preloaded checkpoints; cold-start latency is neglected [13]. Each episode comprises requests, with request originating from PAD and belonging to task suite . The controller selects an execution location and reasoning threshold for each request, fixed throughout the task. Requests are served sequentially, with only one active UL transmission at a time [14]. We use synchronous VLA inference: each cycle begins after the previous action chunk is executed, avoiding potentially stale observations in asynchronous execution [3].

Figure 1. Overview of the proposed WEALR framework.
System Model
Latent Reasoning for VLAs
The VLA processes visual observations and language instructions to generate action chunks, each comprising a sequence of actions (e.g., five). During each inference cycle, recurrent updates refine the hidden representations until the predicted action chunk stabilizes, as illustrated in Figure 1 (b) and (c). Following RD-VLA, convergence is assessed using the squared Euclidean distance between consecutive predicted action chunks as a proxy for the Kullback-Leibler (KL) divergence between their corresponding action distributions:
where denotes the predicted action chunk at recurrent update , indexes inference cycles, and is the maximum reasoning depth. Thus, a single threshold governs each task, while the realized depth varies across observations. Unlike fixed-depth reasoning [14], a task-level threshold adapts depth across observations, potentially reducing recurrent updates while maintaining task success [11].
An offline dataset profiles task success, inference latency, energy, and recurrent-update counts across task suites and candidate thresholds, as summarized in Figure 2. Increasing the convergence threshold reduces the mean recurrent iterations per inference across all suites, but generally lowers task success, with the largest degradation observed in LIBERO-10, which involves long-horizon tasks [17]. These tradeoffs motivate adaptive threshold selection and provide the profiles used by the resource-management environment.
Figure 2. VLA task performance versus inference cost
Communication Model
We consider an orthogonal frequency division multiplexing (OFDM) UL channel in which single-antenna PADs communicate with a BS equipped with antennas. Indoor channels are generated using DeepMIMO by mapping each PAD’s position to the nearest channel location, capturing geometry-dependent blockage and LOS/NLOS multipath propagation [7]. PAD mobility follows the random waypoint model. For request , the channel of PAD on subcarrier is . To represent bandwidth availability under exogenous traffic at the BS, is sampled uniformly from a predefined set for each request and mapped to a central contiguous subcarrier set . The total transmit power is allocated equally across the active subcarriers. With maximum-ratio combining, the UL signal-to-noise ratio (SNR) and achievable rate are:
where is the active subcarrier set, , is the subcarrier spacing, is the noise spectral density, and is the linear receiver noise figure. Only the requesting PAD transmits, excluding interference among the modeled PADs. Each inference cycle uploads two uncompressed RGB images ( Mbit), neglecting prompt and proprioceptive payloads. The upload latency and transmit energy for inference cycle of request are:
Downlink transmission delay and receive energy are neglected due to the small returned action-chunk payload (e.g., 140 bytes). We assume that the PAD remains stationary during each task execution and the UL rate remains fixed throughout each task.
Computation Model
Let and denote local and edge execution, respectively. The inference latencies and account for the respective hardware capabilities and selected convergence threshold. The total chunk latency is
To account for the server’s finite computation capacity shared among PADs, edge workload is measured in reasoning iterations [14]. PAD energy and edge reasoning consumption for task are
where is the number of inference chunks, and denote PAD inference and UL transmission energy, respectively, and is the reasoning iteration count. Locomotion energy is excluded, since PAD motion is independent of the inference execution location. Cumulative energy for PAD and shared edge consumption are and .
A task violates the SLO if more than a fraction of its chunks exceed the deadline :
Problem Formulation and CMDP Framework
Problem Statement
We maximize expected task success subject to PAD energy, edge reasoning, and SLO constraints:
s.t.
where if the PAD successfully completes the intended task. Here, is PAD ’s total energy consumption, is the total edge recurrence count, and and are their respective budgets. Constraints bound expected episode costs; episode feasibility is assessed using realized costs.
CMDP Formulation
The underlying system is modeled as a finite-horizon CMDP with decision steps, one per task request. The policy specifies action probabilities given a nine-dimensional observation. Since the individual remaining energies of non-requesting PADs are omitted, this observation provides an approximate state representation.
Observation Space
Before serving request , the agent observes
where is a four-dimensional one-hot task-suite encoding (Spatial, Object, Goal, or LIBERO-10). Here, denotes the remaining energy of requesting PAD , while the last term represents the maximum consumed-energy ratio across all PADs. In addition, is the remaining edge reasoning budget, where denotes the edge recurrence count for request .
Action Space
Each action selects an execution location and a convergence threshold, both fixed throughout the task:
yielding 12 possible actions.
Reward and Costs
Upon task completion, the reward is and the costs are
These costs measure PAD energy, edge reasoning, and SLO violations, respectively. With and , their episode sums recover the quantities in (10a)–(10c).
CDRL-Based Adaptive Latent Reasoning
TRPO-Lagrangian
We adopt TRPO-Lag to solve (10), combining trust-region policy updates with learned constraint penalties [15, 18]. Given observation , the actor samples from the 12 available actions. The selected configuration is maintained throughout the task, after which success and resource costs are recorded.
Let index the constraints, with cumulative costs and limits . The policy maximizes the Lagrangian objective:
where is an adaptive multiplier and scales the corresponding cost. Unlike manual reward shaping with fixed weights, this formulation dynamically adapts the penalties to satisfy the constraints directly [16].
To optimize (10), we employ an alternating primal-dual scheme. In the primal step, a reward critic and three cost critics estimate future success and cumulative costs from . Using generalized advantage estimation (GAE), we form
This advantage is normalized over the rollout, yielding . With the multipliers fixed, the actor update solves the sampled trust-region problem
Here, denotes the rollout average. The KL bound limits changes in the probabilities of offloading and threshold choices between updates. We approximately solve (17) using damped conjugate gradients and backtracking, accepting only steps that improve the surrogate objective and satisfy the KL bound. The critics are fitted to their respective GAE return targets. In the dual step, after each window of completed episodes, the multipliers are updated using a projected multiplier update:
where is the mean episode cost, is the constraint-specific scaling factor, is the dual learning rate, and . Thus, exceeding a resource or SLO limit increases its penalty, while operating below the limit reduces it.
Simulation Scenarios and Results
Experimental Settings and Baselines
Experimental Settings
We consider PADs served by an indoor BS with a co-located edge server, with three PAD locations sampled from LOS regions and two from NLOS regions. Each episode comprises task requests. Methods are trained across three random seeds and evaluated deterministically over 100 held-out episodes. Table I lists the simulation parameters and TRPO-Lag hyperparameters. The offline dataset described in § 2-A1 comprises 2,400 RD-VLA rollouts across four LIBERO task suites, ten tasks per suite, six convergence thresholds, and ten rollouts per task–threshold configuration. Before training, rollouts are partitioned into fixed training (80%) and test (20%) subsets. Inference latency and GPU power are profiled on an RTX 4090, with power sampled through NVML and energy computed by trapezoidal integration. To emulate Jetson Thor execution, profiled latency and energy are scaled by empirical factors and , respectively [3, 4, 19].
We evaluate WEALR using an offline, profile-driven simulator. At each decision step, a task request is sampled, and the controller selects its execution location and reasoning threshold based on observation . A matching RD-VLA rollout is then sampled from the corresponding dataset split. Its recorded task outcome and action-chunk-level statistics are combined with hardware cost scaling and the simulated wireless link to determine task success, end-to-end latency, PAD energy consumption, and edge reasoning cost.
| Simulation | Value | TRPO-Lag | Value |
| DeepMIMO scenario | I3_2p4 | Hidden layers | [64, 64] |
| 2.4 GHz | Training steps | ||
| 32 | Rollout length | 1024 | |
| {1, 2, 5} MHz | Critic batch size | 64 | |
| 20 dBm | Critic epochs | 10 | |
| 7 dB | |||
| (PAD Velocity) | [0.5, 1.5] m/s | 1 | |
| , , | 0.05, 0.1, 200ms | 0.95 | |
| 1.69 | 0.01 | ||
| 0.535 | CG iterations | 10 | |
| 2.27 kJ | 0.05 | ||
| 40 |
Table 1. TRPO-Lag and Simulation Parameters.
Benchmark Methods
We compare TRPO-Lag with: Task-aware heuristic (TAH), which selects a fixed action per task suite using training-data estimates; Task and rate-aware heuristic (TRAH), which additionally considers the current achievable UL rate; Random, which uniformly samples the twelve actions; PPO and TRPO, which maximize task success without constraints [15, 18]; CPO, which incorporates all three constraints into trust-region policy updates [15]; and PPO-Lag, which handles the same constraints using adaptive Lagrange multipliers [15, 16].
Results
In terms of CDRL performance, Figure 3 compares learning performance and normalized CMDP constraint costs. All methods improve task success during training, reaching approximately 92-94 successful tasks per episode, equivalent to the average episode reward. However, maximizing success alone does not ensure feasible operation. PPO and TRPO maintain SLO violation rates approximately four times the prescribed limit, while their edge reasoning costs approach or exceed the budget during training. In contrast, TRPO-Lag and PPO-Lag reduce SLO violations below the limit while maintaining comparable task success. TRPO-Lag stabilizes at approximately 0.8 of the maximum PAD energy budget and 0.7 of the edge reasoning budget. CPO also controls resource consumption, but its SLO violation rate remains close to the constraint boundary. Figure 4 evaluates task success under the operating constraints on the test data.
Figure 3. Training Performance and Normalized Constraint Costs
Figure 4. Task Success and Constraint-Aware Performance
Although most methods achieve task success rates near 97%, their feasibility-weighted success differs substantially. TRPO-Lag achieves approximately 90%, compared with 85% for PPO-Lag, 73% for CPO, and 64% for TRAH, corresponding to relative improvements of approximately 6%, 23%, and 41%, respectively. This favorable balance between task success and feasibility is consistent with the aggregate benchmark findings in [15] and can be partly attributed to the combination of KL-constrained policy updates and adaptive constraint penalties. Unconstrained PPO achieves only approximately 4% feasibility-weighted success, while TRPO, TAH, and Random yield approximately zero. This gap highlights the importance of explicitly accounting for operating constraints when optimizing task success. In terms of offloading policy, Figure 5 illustrates adaptation to wireless conditions. TRPO-Lag rarely offloads below approximately 25 Mbps, where communication delay limits the benefit of edge execution. Its offloading probability increases with achievable UL rate, reaching approximately 88% at high rates, demonstrating channel-aware inference decisions. Although PPO-Lag and CPO offload more frequently in this region, TRPO-Lag achieves the lowest overall SLO violation rate, approximately 2.2%, compared with approximately 3% for PPO-Lag and 3.4% for CPO. PPO and TRPO offload frequently even at low rates and incur violation rates near 19%. Figure 6 explains the learned suite-dependent decisions through
Figure 5. Offloading Behavior and SLO violations
Figure 6. TRPO-Lag action distribution by task suite
the offline characteristics in Figure 2. TRPO-Lag predominantly selects for Goal and Object, preserving high task success while requiring fewer recurrent iterations than stricter thresholds. For LIBERO-10, which comprises longer-horizon tasks, it selects in approximately 82% of requests, consistent with the substantial success degradation observed at larger thresholds. Location selection complements these decisions: Spatial and LIBERO-10 are offloaded in 63.0% and 67.9% of requests, whereas Goal and Object are primarily executed locally. These results indicate that TRPO-Lag also learns suite-dependent operating points that preserve task success while reducing inference latency and managing shared edge resources.
Conclusion
In conclusion, this paper proposed the WEALR framework for adaptive latent reasoning of VLA models in wireless edge-assisted robotic networks. By formulating the joint optimization of execution location and latent convergence threshold as a CMDP, our Lagrangian-CDRL approach dynamically balances task success against PAD energy budgets, shared edge-reasoning limits, and latency SLO constraints under time-varying wireless conditions. Evaluations across LIBERO manipulation tasks with profiled hardware costs and realistic channel models demonstrate that the proposed framework effectively satisfies resource constraints while achieving up to a 23% gain in feasibility-weighted task success rate over baselines. Future work may explore concurrent robotic task requests with joint wireless and reasoning optimization.
Acknowledgment
This work has been supported by the NSERC Canada Research Chairs Program.
References
- [1]C. Wang et al., “Phyai: Real-time physical ai at the edge, scalable rollouts in the cloud,” 2026. [Online]. Available: https://arxiv.org/abs/2608.03682
- [2]M. J. Kim et al., “OpenVLA: An open-source vision-language-action model,” in Proc. CoRL, ser. PMLR, vol. 270, 06–09 Nov 2025, pp. 2679–2713.arxiv.org/abs/2406.09246
- [3]W. Jiang et al., “How fast can i run my vla? demystifying vla inference performance with vla-perf,” 2026. [Online]. Available: https://arxiv.org/abs/2602.18397
- [4]K. Zhou et al., “Characterizing vision-language-action models across XPUs: Constraints and acceleration for on-robot deployment,” in Fortythird ICML, 2026.
- [5]H. Luo et al., “Ai reasoning for wireless communications and networking: A survey and perspectives,” ACM Comput. Surv., vol. 58, no. 13, Jun. 2026.arxiv.org/abs/2509.09193
- [6]L. Kundu et al., “Ai-ran: Transforming ran with ai-driven computing infrastructure,” IEEE Commun. Mag., vol. 64, no. 1, pp. 168–174, 2026.arxiv.org/abs/2501.09007
- [7]A. Alkhateeb, “Deepmimo: A generic deep learning dataset for millimeter wave and massive mimo applications,” 2019. [Online]. Available: https://arxiv.org/abs/1902.06435
- [8]M. Farzanullah et al., “Wireless multimodal foundation model (wmfm): Integrating vision and communication modalities for 6g isac systems,” 2025. [Online]. Available: https://arxiv.org/abs/2512.23897
- [9]F. Temiz, S. Salehi, and M. Erol-Kantarci, “Wisdom: Wireless sparse decision transformer with mixture-of-experts for multi-task mobile network optimization,” 2026. [Online]. Available: https://arxiv.org/abs/2609.00284
- [10]J. Geiping et al., “Scaling up test-time compute with latent reasoning: A recurrent depth approach,” in The Thirty-ninth NeurIPS, 2025.arxiv.org/abs/2502.05171
- [11]Y. Tur et al., “Recurrent-depth vla: Implicit test-time compute scaling of vision-language-action models via latent iterative reasoning,” 2026. [Online]. Available: https://arxiv.org/abs/2602.07845
- [12]L. Yuan et al., “Device-Cloud Collaborative LLM Inference with Multi-Modal, Multi-Task, Multi-Turn Conversations ,” IEEE Trans. Netw., no. 01, pp. 1–1, Jul. 2026.
- [13]A. Zhou et al., “Ecovla: Energy-efficient device-edge co-inference for vision-language-action models under real-time constraints,” in Advanced Parallel Processing Technologies, 2027, pp. 625–631.arxiv.org/abs/2608.15502
- [14]G. Liu, H. Du, and K. Huang, “MORES: Mobile reasoning-as-a-service via distributed LLM inference-time scaling,” arXiv preprint arXiv:2607.08116, 2026.arxiv.org/abs/2607.08116
- [15]A. Ray, J. Achiam, and D. Amodei, “Benchmarking safe exploration in deep reinforcement learning,” OpenAI, Nov. 2019. [Online]. Available: https://openai.com/index/benchmarking-safe-exploration-in-deep-reinforcement-learning/
- [16]S. Khairy et al., “Constrained deep reinforcement learning for energy sustainable multi-uav based random access iot networks with noma,” IEEE JSAC, vol. 39, no. 4, pp. 1101–1115, 2021.DOI
- [17]B. Liu et al., “Libero: Benchmarking knowledge transfer for lifelong robot learning,” arXiv preprint arXiv:2306.03310, 2023.arxiv.org/abs/2306.03310
- [18]J. Schulman et al., “Trust region policy optimization,” in Proc. ICML, 2015, pp. 1889–1897.arxiv.org/abs/1502.05477
- [19]K. Soltani et al., “Energy-efficient split learning for resource-constrained environments: A smart farming solution,” in 2025 IEEE MASS, 2025, pp. 111–119.arxiv.org/abs/2509.02549




