acceptodds
Under review as a conference paper at ICLR 2027

ChainVLA: Chaining Vision-Language-Action Queries through a Unified Execution State for Long-Horizon Manipulation

Abstract

Humans perform long-horizon manipulation by retaining knowledge of what earlier actions have established while continuously adapting the motion underway. By contrast, action-chunked vision-language-action (VLA) policies repeatedly replan from the current input at each query. Existing methods preserve either long-term task evidence through memory or short-term motion through action reuse and ensembling, leaving the cross-query handoff incomplete. We introduce ChainVLA, a 1.2B-parameter VLA policy that chains successive queries through a single, revisable execution state. Progress Context combines a recurrent Working State with sparse event memory to carry observation-derived task progress, while Motion Tail feeds the preceding prediction's unexecuted continuation into state construction and action generation. Together, the two components condition a decoder that regenerates each action horizon under the latest observation, allowing the carried state to guide the next prediction without fixing it. ChainVLA reaches 62.8% average success on RMBench and 98.8% across four LIBERO suites, while removing Motion Tail or Progress Context reduces RMBench success to 11.2% and 3.0%, respectively. These asymmetric ablations are consistent with motion continuity helping preserve the observation stream from which task progress is inferred.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Long-Horizon Manipulation, Robot Learning, Execution State, Memory

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_CVcGOpenReview ↗
Rating
3(+1)
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ChainVLA for long-horizon manipulation. At each query, the policy uses memory of earlier task progress and the unexecuted actions from its previous prediction. The goal is to keep track of what has happened while maintaining continuity between action chunks. The method reports strong results on RMBench and LIBERO, with ablations of the memory components, action continuation, and stage supervision.

Critical strengths
  1. The paper addresses a practical problem: long tasks require both memory of earlier events and continuity between action chunks. ChainVLA brings these together and reports promising results on RMBench and LIBERO. Removing either component leads to a large drop in success.

  2. The smoothing controls are useful. They show that reducing command discontinuities alone does not recover task success. The separate tests of memory retrieval and the two ways of using the previous action tail also help explain the gains.

Critical weaknesses
  1. The main RMBench result uses extra stage supervision. Removing it and switching to periodic memory writes lowers success from 62.8% to 53.4%, but this changes two things at once. It is also unclear whether external baselines receive equivalent supervision, making the source of the reported advantage difficult to assess.

  2. The paper includes matched internal simplifications, such as removing tail tokens, but lacks a comparison with an established learned action-continuation method under the same memory and supervision setup. The ablations establish the value of the tested components, but do not show whether the proposed integration outperforms an alternative learned continuation design. There are also no real-robot experiments. Real-robot testing is especially relevant here because the method aims to maintain motion across action chunks. Smaller command jumps in simulation do not guarantee smoother movement on a robot. Delays and tracking errors may make the previous action tail less useful, and the current experiments do not test this.

Questions for authors
  1. How are the stage labels obtained? Do they require manual annotation, or are they derived from the simulator? Do any of the comparison methods receive equivalent supervision? Can you separate the benefit of stage supervision from the choice of when to write memories?

  2. Do you have a comparison that keeps the same memory module and supervision, but uses an existing learned action-continuation method? This would help me judge what the proposed integration adds.

Reviewer_vmpEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ChainVLA, a VLA framework that maintains a unified execution state across consecutive action chunks for long-horizon manipulation. The method combines Progress Context, which tracks task progress through recurrent working states and sparse event memory, with Motion Tail, which reuses unexecuted actions from previous predictions to improve motion continuity. The authors evaluate the method on RMBench and LIBERO, achieving 62.8% and 98.8% average success rates, respectively.

Critical strengths

The paper is generally well written and easy to follow. I think addressing task progress and motion continuity together is an interesting and practically relevant direction for VLA research. The improvements on RMBench are substantial, and the authors provide detailed ablation studies to analyze the contribution of different components.

Critical weaknesses

My main concern is the lack of real-robot experiments, especially since motion continuity is one of the main motivations of this work. Simulation can oversimplify many practical challenges in robotic control, such as inference latency, asynchronous execution, and controller tracking errors, making it difficult to assess whether the proposed method would actually work in real-world VLA deployment.

Another major concern is the insufficient comparison with existing action continuation methods. Motion Tail is closely related to previous work on RTC, training-time RTC, and native action continuation, but the experiments mainly compare against relatively simple temporal ensembling and linear continuation baselines. I do not think these comparisons are sufficient to establish the novelty or advantages of the proposed approach.

Finally, the external baseline comparisons are not fully controlled, and all internally trained experiments use only a single random seed. Although the ablation results show large performance differences, it remains unclear how much of the improvement comes from the proposed unified execution state rather than combining existing memory and action continuation techniques.

Questions for authors
  1. How does Motion Tail differ from existing action continuation methods? Approaches such as RTC, training-time RTC, and native continuation have already explored conditioning new action predictions on previous actions. It would be helpful to clarify the methodological differences and provide stronger comparisons with these approaches.

  2. How would the proposed method perform under realistic inference latency and asynchronous execution? The current experiments use a fixed action horizon and execution length without varying control frequency or inference delay. Since Motion Tail relies on the unexecuted portion of the previous action chunk, I think it is important to evaluate how robust the method is when the actual execution timing differs from the assumed setting.

  3. Can the authors provide real-robot experiments or more realistic deployment evaluations? I understand that real-world experiments require additional resources, and I do not think every VLA paper necessarily needs them. However, since motion continuity is a central contribution of this work, I believe real-world validation is particularly important to demonstrate its practical value.

  4. Can the authors provide stronger controlled baselines to isolate the contribution of the unified execution state? For example, combining an existing memory-based VLA with a native action continuation method would be a more convincing comparison. The current ablations mainly demonstrate that removing components from ChainVLA reduces performance, but do not fully establish the advantage of the proposed design over alternative approaches.

Reviewer_whv2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces ChainVLA, a vision-language-action model for long-horizon robotic manipulation. The method maintains a unified execution state containing Progress Context, which stores information about task progress, and Motion Tail, which preserves the unexecuted portion of the previous action prediction. Both components influence the next action generation step, allowing the model to continue a task while adapting to new observations. The method achieves 62.8% average success on RMBench and 98.8% across the four LIBERO suites also a nice ablation of memory and motion components.

Critical strengths
  1. This paper has good motivation and clear need of temporal memory of what has happend and it maintains consistent motion between consecutive action predictions in long horizon robotic tasks.
  2. Very good results and numbers such as 62.8% on five RMBench tasks and also compared with strong baseline. LIBERO performance is also strong across several manipulation settings.
  3. very useful ablations to know what actually contributes such as removing motion tail, progress context and so on .
Critical weaknesses
  1. The previous work such as memoryVLA already has temporal dependent robotic manipulaton and RTC and legato also introduces action continuation across chunks, ChainVLA combines progress memory and unfinished action continuation through a shared execution. The results shows great performance but I wonder if you combine existing memory and continuation mechanism.
  2. Trained configs uses single training seed and kind of make it hard to asses the stability of numbers especially given the large performance changes in the pathway ablations.
  3. Main table includes source reported numbers from models with different configs , comparisons are useful for context but it does not necessarily establish the improvement attributable to this architecture under this controlled training protocol.
  4. the paper implies that motion continuity improved the quality of observations for memory system and all numbers back it up consistently but in my opinion it does not directly measure whether memory retrieval or stage estimation becomes more accurate.
Questions for authors

you can infer from weaknesses but particularly I am interested in this one:

What is the key technical advantage of the unified execution state over combining an existing memory-based VLA method with an action-continuation mechanism? (memoryvla, RTC and legato, basically weakness 1)

2 comments
Reviewer_whv2OpenReview ↗

Thank you Authors.

Thank you for the detailed response and additional experiments. The controlled RTC comparison, direct stage and memory analysis, and additional seeds help address my concerns, although the seed study and real-robot evaluation are limited in scope; I will maintain my score since it justifies the current state of the paper and rating is positive so I am with you.

Thank you again and Good luck.

AuthorsOpenReview ↗

Thank you for your support

Thank you very much for your support and for the constructive feedback throughout the review. Your question on what the unified execution state adds beyond combining existing memory and continuation methods led to the controlled comparisons that are now part of the revised Table 2. Following your suggestion, the final version will extend the seed study to more configurations and expand the real-robot evaluation to more long-horizon tasks, so that both cover a broader range of settings.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.