acceptodds
Under review as a conference paper at ICLR 2027

Requery Is Not Free: Cross-Query Coupling in Generative Robot Control

Abstract

Action-chunked policies have become a widely adopted and effective approach to robot control. These policies repeatedly generate new action chunks from fresh observations, allowing them to adapt during execution. Yet each new chunk must continue a trajectory shaped by earlier actions. Fresh observations update what the policy sees, but may not reveal how that trajectory was reached. A new chunk may therefore fit the current observation yet be incompatible with the execution already underway. We formalize this dependence as cross-query coupling and prove that identical per-query chunk distributions can yield exponentially different trajectory success. In learned policies, we find that a chunk's value depends on the physical trajectory it continues, with particularly pronounced effects in some tasks and execution stages. This diagnosis motivates Stateful Requery, which combines fresh observations with a compact record of realized execution. The record preserves what earlier actions actually changed, giving each new generation context for continuing the current trajectory. The method improves control across all tested backbone–generator combinations on Meta-World MT50 and LIBERO-90. Analysis shows how the record guides branch-specific action adjustments. It remains useful even when the policy requeries after every action. These findings identify requery as a trajectory-composition problem, where effective control requires action chunks that are not merely plausible but appropriate for the execution they continue.

Primary area:
applications to robotics, autonomy, planning
Keywords:
generative robot control, action chunking, cross-query coupling, execution history, imitation learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_oX4MOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper considers the setting of generative robot policies using action chunking, where policies execute a prefix of predicted actions in the environment before predicting a new chunk. The paper states that a problem exists in this setting, as policies may only observe the current observation, and thus, may not have sufficient history to predict actions. However, instead of casting this as a standard setting for history, the paper introduces the terminology of a “cross-query coupling”, which seems to describe a joint distribution between the policies history and the predicted action chunks. Later, the paper discusses that the policy should be aware of history, and then proposes “Stateful-Requery”, which uses a recurrent state m to capture history over proprioceptive information and elapsed time. The paper evaluates adding their method to policies on Libero-90 and MT50 as additions to policies, but does not compare to prior work improving policies for partial observability/history, although related work is partially discussed in the paper.

Critical strengths

Strengths:

  • The paper studies the problem of integrating history into policies, which is an active area of research.
Critical weaknesses

Weaknesses:

  • The writing of the paper is very unclear. This is across the problem being studied, the gaps of prior work, the proposed contribution.

  • The proposed methods appears to be a recurrent state baseline and is not compared to prior work

  • The paper introduces a large amount of unclear jargon, which is not situated within the literature, for example:

    “We call this repeated generative handoff requery”,

“diagnosis of branch-relative continuation value. We turn the criterion into a fixed-bank test, exposing branch-dependent value”

“One query, physical commitment, and return to the interface obey”

The use of the word fiber: “Every candidate gap is unchanged across histories in the fiber”, “they form the interface fiber”

Overall from a presentation, and clarity point of view, I would recommend for the writing of the paper to be completely revised, and simplified. Technically, the paper also lacks both theoretical and empirical study of existing methods for incorporating memory.

Questions for authors
  • Sections 2-3 are very unclear, the paper claims to contribute a “decision-theoretic foundation for cross-query coupling”, and “branch-relative continuation value”; what is established here? In the experiments/method section, it is unclear how this discussion relates to the proposed mechanism, where a policy trained with behavior cloning has a recurrent history.
Reviewer_o4i8OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents a method for robot policy learning so that the policy, like diffusion policy or flow matching based, can condition on not only current observations or past few time steps, but also a learnable encoding that takes previously executed actions and states into consideration.

Critical strengths

This work is generally clearly written and easy to understand.

The proposed memory encoding neural network is interesting that it can efficiently maintain history information without increasing budget. To the best of my knowledge, it seems that this comes from the advantage of Kimi Delta Attention (KDA).

Critical weaknesses

This paper is targeting a classical problem in policy learning or imitation learning where the current state is partially observable, and the agent does not have full state representation that can encompass all historic information. Unlike MDP, in this POMDP setting one has to learn aggregated state/action representation so that the agent know the history or stages. This problem setting has received tons of attention in the past decade and there are many classical algorithms in the literature. This work did not bring new innovative recipe to this problem even it is designed and tested with modern diffusion/flow matching policy.

All the experiments are executed in simulated environment. There is no demonstration on real robot platform, which is essential for a robot learning work.

Questions for authors

Please see the weaknesses part.

Reviewer_3C2GOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that when an action-chunked generative policy "requeries", each new chunk must fit the physical branch created by earlier execution, not only the current observation. It formalizes this as "cross-query coupling" and gives a criterion for which history distinctions matter. The proposed Stateful Requery conditions generation on a recurrent record of realized transitions and improves success on Meta-World MT50 and LIBERO-90.

Critical strengths
  1. The paper reframes requery as a problem of continuing a trajectory, not just predicting good chunks. The key insight is that a chunk’s value depends on the physical branch it continues, which per-query imitation ignores. This view is useful for anyone building action-chunked policies.
  2. The method follows directly from this insight and is simple. Stateful Requery keeps a memory that only executed transitions can write, never unexecuted plans, and the generator reads it alongside fresh observations. The design can work with various backbone or generator, and its benefit does not depend on the specific memory architecture.
Critical weaknesses
  1. I found the paper not easy to follow. It introduces many new terms (interface fiber, coupling diameter, counterbranch, etc.) with little intuition, and key metrics are defined only in App. B.1. In addition, many notations are undefined, such as in Section 4, what actually is UϕU_\phi and what does it do? What is Yq∗Y_q^* in line 319? I encourage the authors make the method section self-contained with clear definition.
  2. The author proposed method starts from Section 4 and it is very brief. For example, for Section 4.2, I would encourage the authors to describe how KDA work and how it can achieve the "fixed-capacity realization" claimed in the paper. This also introduce another concern: seems like the proposed method simply apply a KDA on execution history, and use the resulting features mqm_q to condition the action policy. While I appreciate simple methodology, the contribution here seems weak.
Questions for authors
  1. Why is observation-only Flow so weak on LIBERO, as in many policy learning papers, success rate on Libero-90 is much higher [1][2]?So I wonder why the baseline results in Table 1 is lower than many existing papers.

  2. A follow-up question is that why the Flow model and diffusion model have such a big gap on Libero-90. In my understanding, they should perform similarly, for example, current VLA and WAM also use flow-matching as their generative backbone, and have shown strong performance. Could you explain the performance gap in your experiments?

[1] Flowing With Purpose: Latent Action Guided Flow Matching Policies For Robotic Manipulation
[2] Efficient Diffusion Transformer Policies with Mixture of Expert Denoisers for Multitask Learning

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.