betaFeedback

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:

  1. We develop WEALR to jointly select VLA execution location and reasoning threshold according to task requirements, wireless conditions, and available resources.

  2. We formulate a CMDP with per-PAD energy, shared edge-reasoning, and SLO constraints, and apply TRPO-Lag to optimize robotic task success.

  3. 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 NN 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 KK requests, with request kk originating from PAD iki_k and belonging to task suite qkq_k. 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].

Overview of the proposed WEALR framework

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:

Δk,m(r)=∥ak,m(r)−ak,m(r−1)∥22,(1)\Delta_{k,m}^{(r)} = \left\lVert\mathbf{a}_{k,m}^{(r)} - \mathbf{a}_{k,m}^{(r-1)} \right\rVert_2^2, \tag*{(1)}
dk,m(τk)=min⁡({r:Δk,m(r)<τk}∪{dmax⁡}).(2)d_{k,m}(\tau_k) = \min\left(\left\{r : \Delta_{k,m}^{(r)} < \tau_k\right\} \cup\left\{d_{\max}\right\}\right). \tag*{(2)}

where ak,m(r)\mathbf{a}_{k,m}^{(r)} denotes the predicted action chunk at recurrent update rr, mm indexes inference cycles, and dmax⁡d_{\max} 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 MM 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 kk, the channel of PAD iki_k on subcarrier nn is hik,k[n]∈CM\mathbf{h}_{i_k,k}[n] \in\mathbb{C}^{M}. To represent bandwidth availability under exogenous traffic at the BS, BkB_k is sampled uniformly from a predefined set B\mathcal{B} for each request and mapped to a central contiguous subcarrier set Lk\mathcal{L}_k. The total transmit power PtxP_{\mathrm{tx}} is allocated equally across the active subcarriers. With maximum-ratio combining, the UL signal-to-noise ratio (SNR) and achievable rate are:

γik,k[n]=Ptx∥hik,k[n]∥2LkN0FNFΔf,(3)\gamma_{i_k,k}[n] = \frac{P_{\mathrm{tx}}\lVert\mathbf{h}_{i_k,k}[n]\rVert^2}{L_k N_0 F_{\mathrm{NF}}\Delta f}, \tag*{(3)}
Rik,kUL=Δf∑n∈Lklog⁡2(1+γik,k[n]),(4)R^{\mathrm{UL}}_{i_k,k} = \Delta f \sum_{n\in\mathcal{L}_k}\log_2\left(1+\gamma_{i_k,k}[n]\right), \tag*{(4)}

where Lk\mathcal{L}_k is the active subcarrier set, Lk=∣Lk∣L_k = \lvert\mathcal{L}_k\rvert, Δf\Delta f is the subcarrier spacing, N0N_0 is the noise spectral density, and FNFF_{\mathrm{NF}} is the linear receiver noise figure. Only the requesting PAD transmits, excluding interference among the modeled PADs. Each inference cycle uploads two uncompressed 224×224224 \times224 RGB images (DUL≈2.41D_{\mathrm{UL}} \approx2.41 Mbit), neglecting prompt and proprioceptive payloads. The upload latency and transmit energy for inference cycle mm of request kk are:

Tk,mUL=DULRik,kUL,Ek,mUL=PtxTk,mUL.(5)T^{\mathrm{UL}}_{k,m} = \frac{D_{\mathrm{UL}}}{R^{\mathrm{UL}}_{i_k,k}}, \qquad E^{\mathrm{UL}}_{k,m}=P_{\mathrm{tx}}T^{\mathrm{UL}}_{k,m}. \tag*{(5)}

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 ℓk=0\ell_k = 0 and ℓk=1\ell_k = 1 denote local and edge execution, respectively. The inference latencies Tk,mloc(τk)T^{\mathrm{loc}}_{k,m}(\tau_k) and Tk,medge(τk)T^{\mathrm{edge}}_{k,m}(\tau_k) account for the respective hardware capabilities and selected convergence threshold. The total chunk latency is

Tk,m=(1−ℓk)Tk,mloc(τk)+ℓk[Tk,mUL+Tk,medge(τk)].(6)T_{k,m}=(1-\ell_k)T^{\mathrm{loc}}_{k,m}(\tau_k)+\ell_k\left[T^{\mathrm{UL}}_{k,m}+T^{\mathrm{edge}}_{k,m}(\tau_k)\right]. \tag*{(6)}

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 kk are

ek=∑m=1Mk[(1−ℓk)Ek,mloc(τk)+ℓkEk,mUL],(7)e_k=\sum_{m=1}^{M_k}\left[(1-\ell_k)E^{\mathrm{loc}}_{k,m}(\tau_k)+\ell_kE^{\mathrm{UL}}_{k,m}\right], \tag*{(7)}
gk=ℓk∑m=1Mkdk,m(τk),(8)g_k=\ell_k\sum_{m=1}^{M_k}d_{k,m}(\tau_k), \tag*{(8)}

where MkM_k is the number of inference chunks, Ek,mloc(τk)E^{\mathrm{loc}}_{k,m}(\tau_k) and Ek,mULE^{\mathrm{UL}}_{k,m} denote PAD inference and UL transmission energy, respectively, and dk,m(τk)d_{k,m}(\tau_k) is the reasoning iteration count. Locomotion energy is excluded, since PAD motion is independent of the inference execution location. Cumulative energy for PAD ii and shared edge consumption are Ei=∑k1{ik=i}ekE_i=\sum_k\mathbf{1}\{i_k=i\}e_k and Gedge=∑kgkG_{\mathrm{edge}}=\sum_k g_k.

A task violates the SLO if more than a fraction η\eta of its chunks exceed the deadline TSLOT_{\mathrm{SLO}}:

zk=1{1Mk∑m=1Mk1{Tk,m>TSLO}>η}.(9)z_k=\mathbf{1}\left\{\frac{1}{M_k}\sum_{m=1}^{M_k}\mathbf{1}\{T_{k,m}>T_{\mathrm{SLO}}\}>\eta\right\}. \tag*{(9)}

Problem Formulation and CMDP Framework

Problem Statement

We maximize expected task success subject to PAD energy, edge reasoning, and SLO constraints:

max⁡πθ Eπθ[∑k=1Kyk](10)\max_{\pi_\theta}\ \mathbb{E}_{\pi_\theta}\left[\sum_{k=1}^{K} y_k\right] \tag*{(10)}

s.t.

Eπθ[max⁡iEiEmax⁡]≤1,(11)\mathbb{E}_{\pi_\theta}\left[\max_i \frac{E_i}{E_{\max}}\right] \le1, \tag*{(11)}
Eπθ[GedgeGmax⁡]≤1,(12)\mathbb{E}_{\pi_\theta}\left[\frac{G_{\mathrm{edge}}}{G_{\max}}\right] \le1, \tag*{(12)}
Eπθ[1K∑k=1Kzk]≤ϵ.(13)\mathbb{E}_{\pi_\theta}\left[\frac{1}{K}\sum_{k=1}^{K}z_k\right] \le\epsilon. \tag*{(13)}

where yk=1y_k=1 if the PAD successfully completes the intended task. Here, EiE_i is PAD ii’s total energy consumption, GedgeG_{\mathrm{edge}} is the total edge recurrence count, and Emax⁡E_{\max} and Gmax⁡G_{\max} 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 KK decision steps, one per task request. The policy πθ(ak∣ok)\pi_\theta(a_k\mid o_k) 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 kk, the agent observes

ok=[qk,Eik,kremEmax⁡,GkremGmax⁡,Rik,kUL,kK,max⁡iEi,kusedEmax⁡],(14)o_k=\left[\mathbf{q}_k,\frac{E^{\mathrm{rem}}_{i_k,k}}{E_{\max}},\frac{G^{\mathrm{rem}}_k}{G_{\max}},R^{\mathrm{UL}}_{i_k,k},\frac{k}{K},\max_{i}\frac{E^{\mathrm{used}}_{i,k}}{E_{\max}}\right], \tag*{(14)}

where qk\mathbf{q}_k is a four-dimensional one-hot task-suite encoding (Spatial, Object, Goal, or LIBERO-10). Here, Eik,krem=Emax⁡−Eik,kusedE^{\mathrm{rem}}_{i_k,k}=E_{\max}-E^{\mathrm{used}}_{i_k,k} denotes the remaining energy of requesting PAD iki_k, while the last term represents the maximum consumed-energy ratio across all PADs. In addition, Gkrem=Gmax⁡−∑t=1k−1gtG^{\mathrm{rem}}_k=G_{\max}-\sum_{t=1}^{k-1}g_t is the remaining edge reasoning budget, where gtg_t denotes the edge recurrence count for request tt.

Action Space

Each action selects an execution location and a convergence threshold, both fixed throughout the task:

ak=(lk,τk),lk∈{0,1},a_k=(l_k,\tau_k),\qquad l_k\in\{0,1\},
τk∈{10−4,2×10−4,5×10−4,10−3,5×10−3,10−2},(15)\tau_k\in\{10^{-4},2\times10^{-4},5\times10^{-4},10^{-3},5\times10^{-3},10^{-2}\}, \tag*{(15)}

yielding 12 possible actions.

Reward and Costs

Upon task completion, the reward is rk=ykr_k=y_k and the costs are

ckE=max⁡iEi,k+1used−max⁡iEi,kusedEmax⁡,(16)c_k^E=\frac{\max_i E^{\mathrm{used}}_{i,k+1}-\max_i E^{\mathrm{used}}_{i,k}}{E_{\max}}, \tag*{(16)}
ckG=gkGmax⁡,ckS=zkK.(17)c_k^G=\frac{g_k}{G_{\max}},\qquad c_k^S=\frac{z_k}{K}. \tag*{(17)}

These costs measure PAD energy, edge reasoning, and SLO violations, respectively. With Ei,1used=0E^{\mathrm{used}}_{i,1}=0 and Ei=Ei,K+1usedE_i=E^{\mathrm{used}}_{i,K+1}, 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 oko_k, the actor samples ak=(lk,τk)a_k=(l_k,\tau_k) from the 12 available actions. The selected configuration is maintained throughout the task, after which success and resource costs are recorded.

Let J={E,G,S}\mathcal{J}=\{\mathrm{E},\mathrm{G},\mathrm{S}\} index the constraints, with cumulative costs Cj=∑k=1KckjC_j=\sum_{k=1}^{K}c_k^j and limits (dE,dG,dS)=(1,1,ϵ)(d_{\mathrm{E}},d_{\mathrm{G}},d_{\mathrm{S}})=(1,1,\epsilon). The policy maximizes the Lagrangian objective:

L(θ,λ)=Eπθ[∑k=1Krk]−∑j∈JλjEπθ[Cj]−djsj,(18)\mathcal{L}(\theta,\boldsymbol{\lambda})=\mathbb{E}_{\pi_\theta}\left[\sum_{k=1}^{K}r_k\right]-\sum_{j\in\mathcal{J}}\lambda_j\frac{\mathbb{E}_{\pi_\theta}[C_j]-d_j}{s_j}, \tag*{(18)}

where λj≥0\lambda_j\ge0 is an adaptive multiplier and sj>0s_j>0 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 oko_k. Using generalized advantage estimation (GAE), we form

A^kLag=A^kR−∑j∈JλjsjA^kj.(19)\widehat{A}^{\mathrm{Lag}}_k=\widehat{A}^{R}_k-\sum_{j\in\mathcal{J}}\frac{\lambda_j}{s_j}\widehat{A}^{j}_k. \tag*{(19)}

This advantage is normalized over the rollout, yielding A~kLag\widetilde{A}^{\mathrm{Lag}}_k. With the multipliers fixed, the actor update solves the sampled trust-region problem

max⁡θ E^[πθ(ak∣ok)πθold(ak∣ok)A~kLag](20)\max_\theta\ \widehat{\mathbb{E}}\left[\frac{\pi_\theta(a_k\mid o_k)}{\pi_{\theta_{\mathrm{old}}}(a_k\mid o_k)}\widetilde{A}^{\mathrm{Lag}}_k\right] \tag*{(20)}
s.t.E^[DKL(πθold(⋅∣ok)∥πθ(⋅∣ok))]≤δKL.\text{s.t.}\quad\widehat{\mathbb{E}}\left[D_{\mathrm{KL}}\left(\pi_{\theta_{\mathrm{old}}}(\cdot\mid o_k)\mathbin{\|}\pi_\theta(\cdot\mid o_k)\right)\right]\le\delta_{\mathrm{KL}}.

Here, E^\widehat{\mathbb{E}} 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 WW completed episodes, the multipliers are updated using a projected multiplier update:

λj←[λj+αλC^j−djsj]+,j∈J,(21)\lambda_j\leftarrow\left[\lambda_j+\alpha_\lambda\frac{\widehat{C}_j-d_j}{s_j}\right]_+,\qquad j\in\mathcal{J}, \tag*{(21)}

where C^j\widehat{C}_j is the mean episode cost, sj=djs_j=d_j is the constraint-specific scaling factor, αλ\alpha_\lambda is the dual learning rate, and [x]+=max⁡(0,x)[x]_+=\max(0,x). 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 N=5N=5 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 K=100K = 100 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 sTs_T and sEs_E, 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 oko_k. 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.

SimulationValueTRPO-LagValue
DeepMIMO scenarioI3_2p4Hidden layers[64, 64]
fcf_c2.4 GHzTraining steps5×1055 \times10^5
MM32Rollout length1024
BB{1, 2, 5} MHzCritic batch size64
PtxP_{\mathrm{tx}}20 dBmCritic epochs10
FNFF_{\mathrm{NF}}7 dBαV\alpha_V3×10−43 \times10^{-4}
vv (PAD Velocity)[0.5, 1.5] m/sγ\gamma1
ϵ\epsilon, η\eta, TSLOT_{\mathrm{SLO}}0.05, 0.1, 200msλGAE\lambda_{\mathrm{GAE}}0.95
sTs_T1.69δKL\delta_{\mathrm{KL}}0.01
sEs_E0.535CG iterations10
Emax⁡E_{\max}2.27 kJαλ\alpha_\lambda0.05
Gmax⁡G_{\max}1.94×1041.94 \times10^4WW40

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 τ=10−3\tau= 10^{-3} 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 τ=5×10−4\tau= 5 \times10^{-4} 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. [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. [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. [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. [4]K. Zhou et al., “Characterizing vision-language-action models across XPUs: Constraints and acceleration for on-robot deployment,” in Fortythird ICML, 2026.
  5. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [18]J. Schulman et al., “Trust region policy optimization,” in Proc. ICML, 2015, pp. 1889–1897.arxiv.org/abs/1502.05477
  19. [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

Paper details

vlnc:W_prk8ub9m
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents