acceptodds
Under review as a conference paper at ICLR 2027

Latency-Tolerant Cloud-Edge Collaborative Vision-Language-Action Models via Emergent Representational Specialization

Abstract

Deploying billion-parameter Vision-Language-Action (VLA) policies on mobile robots creates a systems conflict: semantic reasoning benefits from cloud GPUs, whereas closed-loop control must respond locally despite network delay and jitter. Existing hierarchical and asynchronous policies improve throughput, but their slow-path representations can still arrive stale or require explicit scheduling and delay cues. We introduce CloudEdgeVLA, a cloud-edge policy that treats temporal misalignment as a representation-learning problem. A cloud VLA encodes delayed observations into slowly varying task features, while a lightweight edge head combines the latest available cloud feature with current local vision. During training, current and randomly delayed frames are paired with the same current action target in fresh and stale paths. This objective encourages the cloud representation to preserve task-level information while the edge path supplies state-sensitive corrections, driving emergent specialization. Across four LIBERO suites, CloudEdgeVLA retains 63.8-78.0% success with a 40-step uniform-delay window, whereas VLASH reaches at most 6.4% and the evaluated single-path baselines at most 3.0%. By removing blocking synchronization from the control loop, the design offers a practical route to scalable VLA deployment in which cloud models can grow while edge computation remains lightweight and responsive.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Vision-Language Model, Cloud-Edge Systems, Representation Specialization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_cGqmOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose CloudEdgeVLA, a cloud-edge collaborative policy that achieves robust control against temporal misalignment without requiring delay prompts. This is accomplished by designing a non-blocking asynchronous architecture and a paired-frame dual-path training objective, which enables the cloud VLA model to output delay-resilient task features and fuses them with real-time local vision at the edge. Experiments verify that this method can maintain a high success rate under network delays.

Critical strengths

The framework offloads complex VLA semantic reasoning to the cloud while deploying a lightweight vision-augmented action head at the edge. It successfully balances cloud model scalability with low-latency edge control.

By pairing current and randomly delayed frames under identical action supervision during training, the approach naturally induces cloud representations to become robust against temporal misalignment. This eliminates the need for explicit delay prompts or complex scheduling at inference time, significantly enhancing practicality in fluctuating network environments.

Despite being trained with a maximum delay window of only 20 steps, the method exhibits remarkable generalization during evaluation.

Critical weaknesses

The paper claims that by aligning both stale and fresh features to the same current action, the cloud features will spontaneously retain only task-level information, while state-related information is supplemented by the edge. However, the experiments fail to demonstrate "what" the cloud features actually retain.

The baseline comparisons seem unfair. OpenVLA, OFT, and UniVLA were not trained with delay, yet they were fed delayed inputs during evaluation, whereas CloudEdgeVLA was trained with a delay of dmax=20d_{max}=20 and its edge side always has access to the current frame. It is impossible to determine whether the performance gains stem from the architectural design or from the current-frame input.

The delay is sampled uniformly in terms of steps, with no explanation of the correspondence between steps and milliseconds. Real-world network latency is correlated, bursty, and subject to packet loss and disconnections, whereas the paper only tests a uniform distribution, which is overly idealistic.

Questions for authors

Verify whether the cloud features can decode task identity, goals, and progress, while testing whether instantaneous state information has been successfully stripped away, thereby directly demonstrating representational specialization.

Introduce baseline variants that are trained with the same delay and also have access to the current frame, thereby clearly decoupling whether the performance gains stem from the architectural design or the current-frame input.

Assess performance using real or playback network delay traces (including jitter, packet loss, and disconnections) to replace the overly idealistic uniform distribution assumption.

1 comment

Thank you for your thoughtful suggestions. They help us clarify the learned cloud-edge interface and strengthen the evidence for its robustness.

W1/Q1: What cloud features retain. We probed a frozen checkpoint on 80 LIBERO-Spatial episodes using episode-level splits. A linear task-identity probe was fitted at d=0 and held fixed:

Test delayTask balanced accuracy (%)
089.3±1.9
2083.3±2.2
4076.3±3.4

Mean±SD over three splits. The fixed fresh-trained probe reaches 76.3% at d=40, showing that task conditioning remains accessible across observation ages. Instruction-only accuracy is 100%; this probe measures retained task conditioning.

Layer-wise analysis complements this content probe: under the evaluated inputs, final-layer cosine drift at d=40 is 0.282, versus 0.325/0.537 at layers 16/24. These results support a task-conditioned cloud interface that tolerates staleness, with the local visual stream supplying current scene information for action prediction.

W2/Q2: Separating fresh vision and training. We additionally evaluated SmolVLA [1] and SmolVLA+A2C2 [2] using uniform observation-delay windows (50 trials/task). A2C2's local head corrects stale actions using current RGB.

Success at D=40 (%)Four-suite meanSpatialObject
SmolVLA21.928.85.4
SmolVLA+A2C2—32.05.0
CloudEdgeVLA73.576.475.6

CloudEdgeVLA exceeds this current-observation baseline by 44.4/70.6 points on Spatial/Object. Across four suites, it retains 76.5% of no-delay success, versus 28.7% for SmolVLA.

Paired-frame training supervises fresh and delayed cloud paths toward the same current action alongside identical current vision. Table 2a fixes backbone, data, current vision, and optimization budget on Object at D=10: fresh-only/stale-only/dual training reaches 84.6%/92.4%/94.2%. Thus dual training adds 9.6 points over fresh-only training. Separately, adding local vision raises success from 80.6% to 94.2% under dual training. These controls identify complementary contributions from fresh sensing and paired-frame learning.

W3/Q3: Delay units and network variability. D is the observation-delay window in environment steps. At LIBERO's default 20 Hz control rate, D=10/20/40 corresponds to maximum simulated observation ages of 0.5/1/2 seconds. Delay is resampled at policy updates from the available history.

Uniform random sampling allows abrupt changes in observation age between updates, providing a demanding test of rapidly varying staleness. After training with delays up to 20 steps, CloudEdgeVLA retains 63.8–78.0% success at D=40 across the four LIBERO suites. Our Franka pilot provides complementary physical evidence: at 1000 ms added network RTT, success is 80%/70% for static/displaced-target conditions versus 0%/0% for VLASH (10 trials per condition).

Packet loss and reordering call for communication-level handling; prolonged disconnections require separate fault-detection and safe-stop or local-fallback mechanisms. These deployment safeguards complement our focus on responsive control with delayed cloud information and lie beyond the present scope.

References

[1] Shukor, M., et al. SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics. arXiv:2506.01844, 2025.

[2] Sendai, K., et al. Leave No Observation Behind: Real-Time Correction for VLA Action Chunks. arXiv:2509.23224, 2025.

Reviewer_mD5dOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a robot-control system that combines delayed features from a large cloud model with current visual observations processed locally. This lets the robot continue acting without waiting for cloud responses. Its main innovation is paired-frame training. Both current and delayed cloud observations are supervised toward the same current action, alongside fresh local vision. This encourages the cloud to provide temporally stable task information and the local controller to handle immediate state changes, without requiring explicit delay information. Across four LIBERO simulation suites, it achieves 63.8~78.0% success with delays sampled up to 40 steps, despite training only up to 20 steps. Results include three seeds, component ablations, and representation analyses supporting reduced sensitivity to delay. A small Franka robot pilot achieves 80% static and 70% dynamic success with 1 second of added network delay.

Critical strengths
  1. It addresses how robots can maintain responsive control when cloud computation introduces unpredictable delays.
  2. Paired-frame training directly encourages tolerance to stale cloud features, while current local vision supplies corrective information.
  3. Success remains 63.8~78.0% under delays up to 40 steps, substantially exceeding evaluated baselines and extending beyond the 20-step training window.
  4. Ablations show benefits from both current local vision and delayed-feature training. Representation analysis also shows reduced feature drift and reduced propagation into actions.
  5. The real-robot pilot includes target movement and substantial added latency, providing a useful feasibility check beyond simulation.
Critical weaknesses
  1. The proposed method receives current local vision, while several baselines receive delayed observations. The performance gap therefore mixes advantages from fresh sensing, architecture, and training. Stronger comparisons may give competing methods the same information and adaptation budget.
  2. Reduced feature drift demonstrates temporal stability, but does not prove that cloud features encode task semantics or remain essential. Removing or shuffling cloud features may help test whether the local controller largely solves the task independently.
  3. Simulation uses uniformly sampled observation delays, which does not fully capture correlated latency, packet loss, out-of-order arrivals, or prolonged disconnections.
  4. The physical pilot uses an RTX 5080 workstation for the edge. Performance on constrained robot hardware, control frequency, power consumption, and communication bandwidth remain insufficiently demonstrated.
  5. Real-world evidence covers one pick-and-place task with only 10 trials per condition. Component ablations cover one simulation suite at a moderate delay.
Questions for authors
  1. How does CloudEdgeVLA compare with baselines given the same current local observations and delay-aware training?
  2. Does performance drop when cloud features are removed or shuffled, confirming that the policy meaningfully uses cloud information?
  3. How robust is the method to packet loss, correlated delays, out-of-order responses, and prolonged disconnections?
  4. What control frequency, bandwidth, and power consumption does the system require on resource-constrained edge hardware?
  5. Do the ablation findings and real-robot gains hold across more tasks, higher delays, and additional trials?
1 comment

Thank you for these constructive suggestions. They help us sharpen the contribution and strengthen the evidence for practical cloud-edge control.

W1/Q1: Fresh observations and training. We additionally evaluated SmolVLA [1] and SmolVLA+A2C2 [2] using uniform observation-delay windows (50 trials/task). A2C2's local correction head receives current RGB.

Success at D=40 (%)Four-suite meanSpatialObject
SmolVLA21.928.85.4
SmolVLA+A2C2—32.05.0
CloudEdgeVLA73.576.475.6

CloudEdgeVLA exceeds A2C2 by 44.4/70.6 points on Spatial/Object. Across four suites, it retains 76.5% of no-delay success, versus 28.7% for SmolVLA.

Paired-frame dual-path training is central: fresh and delayed cloud paths share current local vision and the current action target; the delayed path updates both backbone and head. Table 2a controls backbone, data, current vision, and optimization budget on Object at D=10: fresh-only/stale-only/dual training reaches 84.6%/92.4%/94.2%. The 9.6-point fresh-to-dual gain isolates the training contribution with visual freshness held fixed.

W2/Q2: Cloud-feature dependence. Following your suggestion, we zeroed cloud tokens while retaining current local vision (two Spatial tasks; 20 paired initial states each):

Cloud blockFixed d=0Fixed d=20
Intact40/4038/40
Zeroed0/400/40

This confirms the contribution of cloud information alongside fresh vision. Layer-wise analysis shows lower final-layer cosine drift at d=40 (0.282 vs. 0.325/0.537 at layers 16/24) under the evaluated inputs, supporting a stable cloud interface.

W3/Q3: Network conditions. Uniform random delay sampling allows abrupt changes in observation age between policy updates, providing a demanding test of rapidly varying staleness across the full window. CloudEdgeVLA retains 63.8–78.0% success at D=40 after training with delays up to 20 steps. Packet loss and reordering call for communication-level handling, while prolonged disconnections require separate fault-detection and safe-stop or local-fallback mechanisms. These deployment safeguards lie beyond our focus on responsive control under delayed cloud updates.

W4/Q4: Edge hardware. Our laboratory prototype uses an RTX 5080 at the edge and an RTX 4090 for the remote 7B backbone, demonstrating functional cloud-edge control under latency. For deployment, we will consider more constrained edge devices and profile their achievable control frequency, communication bandwidth, and power consumption.

W5/Q5: Broader validation. Ablations, four LIBERO suites, RoboTwin, and the Franka pilot provide complementary evidence. FastWAM-based evaluation covers 50 RoboTwin tasks (2,500 trials per method and delay), reaching 82.32% at D=40 versus 14.72% for StarWAM-MoT across a new benchmark and backbone. On Franka, success at 1000 ms added RTT is 80%/70% for static/displaced-target conditions versus 0%/0% for VLASH (10 trials per condition). We share the reviewer's interest in more complex tasks and larger-scale physical evaluation, an important and challenging direction for the field that we will continue to pursue.

References

[1] Shukor, M., et al. SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics. arXiv:2506.01844, 2025.

[2] Sendai, K., et al. Leave No Observation Behind: Real-Time Correction for VLA Action Chunks. arXiv:2509.23224, 2025.

Reviewer_SsMBOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

CloudEdgeVLA is a framework based paper that divides the task of action planning using a hybrid approach where the reasoning with bigger models on cloud are not used with recent states and the action model is conditioned on representations of reasoning models with older states making action prediction in real time practical.

Critical strengths
  1. Practicality: It is a problem that required necessary attention where waiting for cloud outputs at every step is virtually not possible and makes it unusable, so this is a great step toward a more practical approach to rely on models that are lightweight with only using older representations from a reasoning model as representation affirmation.
  2. Well-ablated: The paper is well ablated and choice of hyperparameters have been justified.
  3. Strong results on Libero: The paper has very strong results on LIBERO compared to the baselines and is miles ahead of the baselines.
Critical weaknesses
  1. LIBERO: The paper mostly claims success on just LIBERO. Results from challenging dataset like RoboTwin would have made the claim stronger.
  2. Long-Range Missing: The paper does not validate the approach against longer range benchmarks and task. It will be better to know how much the result drift as the prediction length increases substantially.
  3. Incomplete description of the representation: The paper does not clearly say from which layer of the network the representation is extracted. This is needed with proper ablation showing why that layer is better than any other layers. Some form of small mechanistic or probing experiment will also be enough to validate the design choice.
Questions for authors

Can you generate more comprehensive results and real-world experiments?

1 comment

Thank you for your thoughtful and constructive feedback, which has helped us identify valuable ways to refine our work and strengthen the paper.

W1: Evaluation beyond LIBERO. We extended evaluation to all 50 RoboTwin tasks using a FastWAM-based implementation of CloudEdgeVLA, with 50 trials per task and delay setting (2,500 trials per method and delay). The overall success (%) is:

MethodD=0D=10D=20D=40Retention
CloudEdge86.9685.8485.1680.8893.0%
StarWAM-MoT85.5278.0044.9614.7217.2%

D is the observation-delay window. At D=40, CloudEdge's success-rate advantage reaches 66.16 percentage points while retaining 93.0% of its no-delay success. Successful application to FastWAM extends the evidence for delay robustness beyond both LIBERO and the original OpenVLA-OFT backbone.

W2: Longer tasks. LIBERO-10 (Long) evaluates longer, multi-stage manipulation, where CloudEdgeVLA achieves 63.8% at D=40 (Table 1). We further stratify all 50 RoboTwin tasks using their official per-task maximum evaluation horizons: short (≤400 steps; 24 tasks), medium (401–800 steps; 18 tasks), and long (>800 steps; 8 tasks). This grouping covers the full benchmark and follows its predefined evaluation limits. The results are shown below (success, %):

GroupMethodD=0D=10D=20D=40Retention
ShortCloudEdge88.9286.5885.4283.1793.5%
ShortStarWAM-MoT87.3381.0052.1718.1720.8%
MediumCloudEdge85.7887.5687.6782.7896.5%
MediumStarWAM-MoT88.2281.5649.7816.5618.8%
LongCloudEdge83.7579.7578.7569.7583.3%
LongStarWAM-MoT74.0061.0012.500.250.3%

Retention is D=40 success relative to D=0. CloudEdge maintains over 80% success on both short and medium tasks at D=40, while achieving 69.75% on long tasks. Its D=40 success-rate advantage is largest on long tasks: 69.50 percentage points across 400 trials per method. On this group, CloudEdge retains 83.3% of its no-delay success, versus 0.3% for StarWAM-MoT. For example, it achieves 98%, 86%, and 74% at D=40 on long tasks blocks_ranking_rgb, stack_blocks_three, and stack_bowls_three, respectively. The full-benchmark coverage and strong long-task retention show that the cloud-edge recipe supports sustained manipulation under delayed cloud updates, extending its benefits across task lengths as well as benchmarks and backbone architectures.

W3: Representation extraction. We use the final transformer-layer outputs at the action-token positions (hidden_states[-1]). The edge head concatenates per-timestep action-token features with current local visual features and predicts actions through a residual MLP. Both training paths supervise this interface toward the same current action.

New diagnostics cover 960 states from 80 LIBERO-Spatial episodes (cosine drift; lower is better):

Layerd=20d=40
160.2740.325
240.4370.537
32 (final)0.1720.282

The final layer is the most temporally stable of the three tested layers; paired 95% bootstrap CIs for both contrasts exclude zero at each delay. This supports its suitability for delayed cloud updates.

Q1: Real-world validation. Our Franka pilot places a toy bear into a box, with static and dynamic conditions. In the dynamic condition, the target is displaced by up to 10 cm during reaching. At 1000 ms of added network RTT, CloudEdgeVLA achieves 80% and 70% success, respectively, versus 0% in both conditions for VLASH (10 trials per condition; Table 2). This provides physical evidence that the cloud-edge design sustains responsive manipulation under substantial latency.

Comments on the paper

2 comments

We thank the reviewers for their thoughtful and constructive feedback. Your suggestions have helped us strengthen the experimental evidence and clarify the presentation. We have updated the manuscript to R1 version as follows.

  1. Table 1 and Figures 3–4. We added SmolVLA baseline using the officially released checkpoint. Table 1 additionally includes SmolVLA+A2C2, which uses current local observations. We revised the abstract and Sections 4.2–4.3 accordingly and clarified the availability of A2C2 checkpoints.
  2. Table 2(b) and Section 5. We added layer-wise representation diagnostics covering 960 states from 80 LIBERO-Spatial episodes. The final layer exhibits the lowest temporal drift among the three tested layers, supporting its suitability for delayed cloud updates.
  3. Section 3.2. We clarified that cloud representations are extracted from the final transformer layer at action-token positions and fused with current local visual features.
  4. Section 4.3. We expanded the discussion of longer, multi-stage manipulation. On LIBERO-Long, CloudEdgeVLA achieves 66.7% success at the 40-step delay window, exceeding the strongest evaluated baseline by 49.1 percentage points, compared with a 4.2-point margin at the 10-step window. This highlights its pronounced advantage on extended task execution under larger observation delays.

These results strengthen the evidence for our central design: paired-frame dual-path training makes cloud representations compatible with delayed updates, while current edge vision grounds responsive actions. At the 40-step delay window, CloudEdgeVLA achieves 75.5% average success across four suites, compared with 21.9% for the strong baseline SmolVLA. Its advantage also extends to the evaluated A2C2 controller with fresh local observations.

We have also added the following LLM Usage Statement to the revised manuscript, before the references:

LLM Usage Statement We used GPT to assist with code development, organization of experimental results, manuscript language polishing, and formatting checks. We also used Qwen, Doubao, and GPT to generate graphical elements for the method illustration. The authors take full responsibility for the final manuscript, code, figures, and scientific conclusions.

We warmly welcome further questions and suggestions. We would be happy to continue this constructive discussion and work with the AC and reviewers to further refine the manuscript and strengthen its quality.

Dear Reviewers,

The authors have submitted their rebuttal. Please review it and share your comments. If anything is unclear or requires clarification, please let the authors know as soon as possible. The author–reviewer discussion period ends on 18 November.

Best regards,

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.