acceptodds
Withdrawn from ICLR 2027

When Foresight Drifts: Progress Aligned Resynchronization for Cloud–Edge Robotic Control

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Deploying billion-parameter vision-language-action (VLA) policies on mobile robots creates a systems tension: semantic reasoning benefits from cloud GPUs, whereas closed-loop control requires responsive local execution. We identify two sources of foresight drift in cloud–edge robotic control. First, uniform temporal prediction allocates representational capacity by elapsed duration rather than semantic significance, creating a pace–progress mismatch that underrepresents critical procedural transitions. Second, physical changes can invalidate predicted continuations even when the semantic stage remains relevant, creating a state-foresight mismatch. We propose Foresight Resync (ForRes), which integrates semantic-target-conditioned cloud prediction, progress-based edge cache addressing, and value-guided asynchronous refresh. In the cloud, a JEPA-based predictor forecasts the next semantic milestone from causal execution history, and a lightweight adapter conditions the world action model on this target while preserving its native prediction budget. On the edge, recurrent addressing matches observed execution to cached semantic anchors, retrieving latent guidance according to procedural progress rather than elapsed time. Target-support gating suppresses incompatible guidance, while an alignment-conditioned critic requests refresh when its expected benefit exceeds querying costs; invalid or exhausted caches trigger refresh directly. Edge policy maintains observation-only closed-loop control while awaiting cloud responses. Evaluated with two cloud backbones and an X-VLA edge policy, ForRes achieves up to 78.73% average success on RoboTwin 2.0, improving over Latent-to-Action by 4.20 percentage points with 31.11% fewer cloud calls, and reaches 76.00% success on real-world dual-arm tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
cloud-edge, WAM;Robotic

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6HZAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Foresight Resync (ForRes), a cloud–edge robotic control framework designed to address foresight drift caused by mismatches between predicted futures and actual robot execution. The method combines semantic milestone-conditioned cloud WAM prediction, progress-aligned retrieval of cached future latents, and validity-aware, value-guided asynchronous refresh. A lightweight edge VLA performs closed-loop control using the selected foresight guidance. Experiments on RoboTwin 2.0 and real-world dual-arm manipulation tasks demonstrate improved task success rates and reduced cloud inference calls compared with existing foresight transfer baselines.

Critical strengths
  1. The proposed technical approach is intuitive and well-motivated.
  2. The experimental results are promising and demonstrate good performance.
Critical weaknesses
  1. The paper proposes a complex framework consisting of multiple interdependent modules, but several key implementation details are missing, including the construction of semantic-event labels, the specific architecture of the JEPA-based predictor, the training procedures of individual modules, and the generation of supervision signals. Furthermore, some methodological descriptions are unclear or inconsistent, making the proposed framework difficult to fully understand and reproduce.
  2. The experimental evaluation primarily focuses on final task success rates and a limited set of component ablations, without directly assessing or thoroughly analyzing key intermediate mechanisms, including milestone prediction, event-weighted foresight generation, progress alignment, cache-validity detection, and continuation-value estimation. Consequently, the current results do not sufficiently demonstrate whether these components achieve their intended functionalities or explain how they contribute to the observed downstream performance improvements.
Questions for authors
  1. Several key details remain unclear, including: (1) the annotation of e(t)e(t) and BtB_t, especially for real-world demonstrations; (2) supervision construction for the progress addressor and target-support gate; (3) the relationship among V-JEPA2, EtarE_{\mathrm{tar}}, and FψF_\psi; (4) the training pipeline, teacher forcing, and cache generation/update procedures; and (5) the implementation of the Latent-to-Language and Latent-to-Action baselines. Could the authors clarify these details to ensure reproducibility and fair comparisons?
  2. Several inconsistencies are present: (1) the method replaces time embeddings with progress coordinates, while the experiments claim to preserve native temporal sampling; (2) Section 5.2 specifies Motus + X-VLA, whereas Table 3 indicates LingBot-VA + X-VLA; and (3) Table 2 reports 20 physical trials per task, yet several success rates cannot be obtained from 20 binary trials. Could the authors reconcile these inconsistencies and clarify the actual evaluation protocols?
  3. The paper lacks direct evaluations of progress-alignment accuracy and invalid-cache detection, as well as an independent ablation of the target-support gate. Could the authors provide these analyses, particularly under execution-speed variations and physical disturbances?
  4. Efficiency improvements are primarily demonstrated through reduced cloud calls, without adequately accounting for additional edge-side computation or overall system latency. Could the authors report the inference overhead of the proposed modules and the end-to-end latency?
  5. The simulation experiments are conducted exclusively on RoboTwin 2.0, limiting evidence of generalizability. Could the authors evaluate the method on another benchmark, such as LIBERO Pro, to assess robustness across different environments and task distributions?
Reviewer_5q31OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents ForRes, a cloud–edge robotic control framework that addresses misalignment between predicted futures and ongoing execution. It combines semantic-milestone-conditioned world-model prediction, progress-based retrieval of cached future latents, a physical-compatibility gate, and a continuation critic that decides when refreshing cloud guidance is worthwhile.

Critical strengths
  • Cached cloud predictions can become misaligned because execution proceeds at a different pace or because the physical scene changes. Treating these as distinct problems is a useful organizing principle.
  • Milestone prediction, progress-based retrieval, validity checking, and asynchronous refresh form a plausible approach to maintaining useful guidance without blocking local control.
  • Relevant evaluation scope. The paper includes two cloud backbones, component ablations, execution disturbances, and physical manipulation experiments...
Critical weaknesses
  • The real-world results require clarification. Table 2 states that evaluation uses 20 matched physical trials, but reports per-task success rates such as 94%, 82%, and 6%. These cannot come directly from binary outcomes over 20 trials (?), whose success rates change in increments of five percentage points. Additional averaging could explain this, but no such protocol is described. Can the authors clarify this? The authors should provide raw success counts, the number of trials and repeated runs, and the aggregation procedure. Moreover, the physical evaluation omits Latent-to-Action, so it does not isolate the benefit of the proposed synchronization mechanisms over the closest collaborative baseline.
  • Section 4.1 replaces fixed-time prediction coordinates with event-weighted progress coordinates ending at a semantic boundary. Section 5.1 nevertheless states that native temporal sampling is preserved, with each latent covering 16 or 8 action steps. These descriptions may refer to different quantities, but their relationship is not explained. The authors should specify how event-based resampling interacts with each backbone’s temporal encoding, latent representation, and prediction horizon.
  • The supervision and training requirements are substantially underdocumented. can the authors add more details on this? The framework requires event boundaries, phase labels, object roles, annotations of generated predictions, compatibility labels, and paired counterfactual simulator rollouts. These are important requirements beyond ordinary action demonstrations.
  • Figure 4 includes AsyncVLA and visual-discrepancy baselines, but their configurations and evaluation protocols are not explained in detail. Figure 4 reports large speedup ratios, but the hardware, absolute latencies, timing boundaries, and precise reference configurations are not adequately specified.
  • The target-support gate is central to handling physical invalidation, but it is not separately ablated.
Questions for authors
  • What raw trial counts and averaging procedure produce the physical success rates in Table 2?
  • How are event-weighted prediction coordinates reconciled with the fixed temporal coverage stated in Section 5.1?
  • Can the authors provide a separate validity-gate ablation and complete hardware, latency, memory, and communication measurements?
Reviewer_xv5oOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces a world action model (WAM)-based framework for cloud-edge collaboration. Cloud inference predicts semantic milestones while generating `guidance' - future latent predictions, which are then cached on the edge. As the cached guidance may become stale, the edge policy determines which guidance to select and whether an update should be requested from the cloud (via a continuation critic), given the expected gain, estimated communication latency, and query cost.

Critical strengths
  • The introduced framework handles practical challenges in cloud-edge collaboration, for instance asynchronous inference. While awaiting cloud responses the edge continues closed-loop control using local observations without any cached cloud-based guidance.

  • The figures and writing are generally clear.

Critical weaknesses

There are several critical concerns around the positioning and analysis of the framework.

  • [W1] The paper could benefit from better positioning, given the large amount of prior cloud-edge collaboration work (for instance, Efficient Edge Inference by Selective Query, Kag et al., ICLR 2023, https://arxiv.org/abs/2608.00569, among others). As very few baselines are considered besides the authors' own, this becomes a critical weakness of the paper in the current form.

  • [W2] The considered Robotwin tasks seem mostly static. A dynamic task, such as objects running on a conveyor belt and the robot has to choose the correct object and grasp it, or arranging objects in a shelf where the shelf is cleared objects are taken out while the task is in progress like in a grocery store, would have been more ideal for a problem formulation to show state mismatch case. That would also highlight performance under more considerable trade-offs, i.e., between cached foresight, current physical state and cloud latency.

  • [W3] The method trains a critic with counterfactual impact, yet it lacks detailed evaluation - for instance, situations where cloud computation latency changes or how often the critic makes correct vs unnecessary cloud switch decisions. Additionally, a clearer formulation of how round-trip delays/refresh cost is calculated or how simulated latency is generated would have been helpful for reproducibility.

  • [W4] Is a learned critic is necessary or is a semantic milestone based refresh enough? This should be clearly evaluated in the paper. I'm also lacking an evaluation whether a target based gate is necessary or a refresh when visual/state error exceeds a threshold gives similar performance to ForRes?

  • [W5] There seem to be several inconsistencies in experiment section. Edge parameters are missing in Table 1 (this is important information for understanding the results). AsyncVLA results are in Fig. 4b and 4c but the numbers are not added in Table 1. Table 3 caption says ForRes full uses LingBot but its numbers match Motus numbers in Table. 1.

Questions for authors

In addition to the aforementioned questions and weaknesses, how does ForRes compare with a simpler baseline that tracks semantic task progress and refreshes guidance when visual prediction error exceeds a threshold?

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

Related papers

Discussion (0)

Sign in to comment.