acceptodds
Withdrawn from ICLR 2027

ROBOT MEMORY IN CONTEXT: SEPARATING STORED CONTENT FROM CONDITIONS OF USE

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

After a cue disappears, a robot must still choose the object it indicated. We ask when a stored representation that can guide this choice remains useful as the scene and control phase change. In a frozen recurrent robot policy, we replace saved states and edit cue components estimated from other source scenes, holding the current test scene fixed within each comparison. On independently sampled scenes, replacing one cue component with another redirects 53/64 choices toward the specified new object, versus 0/64 for token-magnitude-matched random changes: an 82.8-percentage-point gain in targeted redirection. The same edited tensors yield 11/64 matches when inserted two actions later, with the same remaining action budget. What the robot did in between matters: at that later query, an unchanged saved state guides 14/64 choices after the ordinary action prefix and 49/64 when the two preceding arm motions are paused. On jointly new source and test scenes, calibration reduces Brier probability error from .1717 to .0583, a 66.0% reduction against a same-budget rate baseline. A memory can hold the right content yet fail to guide action, so robot memory should be evaluated both for what stored content can direct and for when the controller can use it.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot memory, Causal interventions, Vision-language-action models, Embodied AI

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

This paper studies how stored cue information interacts with a robot controller's current situation. It edits the recurrent state of a frozen muVLA policy, tests cue deletion and conversion, and inserts identical states at different execution boundaries. Action-prefix pauses change whether an edited state redirects behavior. Additional factorial and prediction experiments examine the interaction between stored state and conditions of use.

Critical strengths

S1. The editing controls are strong. Source-excluded component estimates, matched perturbations, deletion and exact restoration link specific edits to specific choices (Section 3; Appendix B). Targeted conversion succeeds in 53/64 cases versus 0/64 for matched noise, with further source-bank and identity tests.

S2. The timing effect is directly tested. Identical tensors and equal continuation budgets show that insertion time matters. The paused-versus-ordinary comparison reaches 49/64 versus 14/64, with no-write controls for each prefix (Section 4; Table 2). Supplied outcome records agree with these totals.

Critical weaknesses

W1. Late failure is not isolated as a memory-use limitation. Pausing changes physical configuration and fresh observations; late insertion also places an earlier recurrent state into a different control situation. Equal horizons and updated-state controls address time and state aging, but do not show whether the late policy can follow any newly supplied target. The evidence establishes conditional effectiveness, without separating memory use from general difficulty redirecting control.

W2. The central chain is confined to one checkpoint and task. The complete editing, timing and recovery sequence is demonstrated for K2 on Shape x Color. Other policies test related but different interventions (Table 1; Appendices D and F), so they do not independently reproduce that sequence. This limits how broadly the proposed account can explain robot memory.

Assessment. I recommend weak rejection (2/4). The intervention effects are credible, but their memory-specific interpretation and reach beyond the principal controller remain insufficiently established.

Questions for authors

Q1 (W1). At the ordinary late boundary, does any existing condition show that the policy can still follow freshly visible target information? If not, which reported result distinguishes failure to use memory from a general inability to redirect the ongoing action?

Q2 (W2). Which additional-policy result tests the same stored state under ordinary and paused prefixes at a matched boundary? Please identify that result, or state precisely which part of the main finding is supported by the different interventions already reported.

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

The paper works on the problem when a memory stored in a frozen robot policy actually influences its behavior after the cue is no longer visible. The authors mainly study a released recurrent VLA using a simulated task. They save the full recurrent state from one episode and transfer it into a matched test episode where the cue is no longer available. By keeping the inputs, simulator state, action history, and random state consistent, they can isolate the effect of the stored memory.

Critical strengths
  1. I think the experimental design is one of the stronger parts of the paper. The authors take care to control factors that could affect the results. For example, paired branches use the same inputs, simulator state, random state, and executed action prefix. They also show that adding back the exact memory component reproduces the original actions and geometry. The random edits are matched in norm for each token, and the confirmation experiments are kept separate from those used during development. The authors also repeat the main experiments with a new source bank and across all nine object identities (Tables 2 and 7).

  2. I also found the difference between what the memory stores and what the policy actually uses particularly interesting. Even when the saved memory can still be decoded correctly in most cases, it only affects the robot's choice in some (13 out of 64) trials when inserted two actions later.

However, when the preceding arm movements are paused, the same memory becomes effective again in 49/64 trials. To me, this is one interesting findings in the paper. It suggests that having information stored in a model's internal state does not necessarily mean the model will use that information to make decisions.

Critical weaknesses

1. Limited generalization beyond the main policy and task.

My main concern is that the central findings rely heavily on a single recurrent µVLA checkpoint (K2) and the relatively simple Shape×Color reaching task. In particular, the timing and pause experiments, which I see as the most interesting contributions, are only studied in detail on this system. The timing effect is also measured over a difference of just two action queries. Although the paper includes additional experiments with K8, Long-hist, GMP, and MemoryVLA, these investigate different questions and do not directly reproduce the main timing and recovery findings.

This is consequential because the paper aims to draw broader conclusions about when robot policies use stored memory. I am not convinced that the evidence is sufficient to establish whether these findings reflect a general property of memory-based robot control or something specific to the K2 checkpoint and its training setup.

2. The timing and pause experiments do not fully distinguish memory use from physical commitment.

The paper shows that the same saved memory becomes much less effective when inserted two actions later, and that pausing the preceding arm movements largely restores its effectiveness. However, I am not convinced that the experiments sufficiently explain why this happens.

One possible explanation is that the robot has already started moving toward a particular target by the later insertion point. In this case, the lower cue-aligned success rate could reflect the difficulty of changing an ongoing movement rather than a failure to use the stored information. The pause intervention changes the robot's physical configuration and subsequent observations, so it does not completely separate these possibilities. Moreover, Appendix C.2 reports that late insertion still changes 36/64 choices relative to no-write, suggesting that the memory continues to influence behavior even when it does not produce the intended choice.

This matters because the distinction between stored content and the conditions under which it guides behavior is the paper's central contribution. While I find the reported effects convincing, I think the interpretation remains less clear than the paper suggests. Together with the limited generalization, this makes me lean toward rejection.

Questions for authors

Apart from the weaknesses mentioned above, I do have some other clarification qusestions and suggestions

Clarification questions

  1. Could the authors provide a more detailed analysis of the robot's movements between query 7 and query 9? Specifically, how far does the arm move, does it move toward the object eventually selected, and how often does the late-insertion choice match the no-write choice? Since Appendix C.2 reports that late insertion still changes 36/64 choices, understanding the remaining cases would help determine whether the timing effect mainly reflects physical commitment.

  2. At what point do the Shape×Color training demonstrations typically begin moving toward the target after the cue disappears? If the policy was mainly trained to use cue information early in the action sequence, could the observed timing effect reflect a learned control-phase preference rather than a more general limitation of memory use?

  3. Can the authors clarify the chance-level success rate in Shape×Color, including the number of candidate objects in each scene? This would help interpret the relatively low no-write baselines (12/96, 7/64, and 10/64).

  4. Regarding the prediction experiments in Section 6, the scene-only model achieves a lower Brier score (0.0520) than the additive source–scene model (0.0583), and context-based source selection achieves 23/32 compared with 25/32 for a global-source baseline. What additional practical benefit does the proposed prediction framework provide beyond identifying scenes where the policy is already likely to succeed?

Suggestions for strengthening the work

  1. I would strongly encourage the authors to repeat the timing and pause experiments on a second recurrent policy or on a task with a longer delay between cue observation and target selection. This would directly address my main concern about the generality of the conclusions.

  2. It would also be useful to include more structured controls for the cue-component interventions, such as components estimated using shuffled cue labels or differences between real recurrent states with the same cue. This could help establish whether the observed effects are specific to the estimated cue representation rather than other correlated properties of the recurrent state.

Reviewer_Bw3TOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper investigates when information stored in a robot policy's memory can guide its behavior. Using frozen policies in simulated manipulation tasks, the authors replace recurrent states across scenes and edit components associated with target cues. They then measure whether the robot selects the corresponding object.

Critical strengths
  • The paper uses a controlled test to determine whether stored cue information actually guides the robot's choice.
  • The paper also tests when that information stops being useful by inserting the same saved state at different times, highlighting the importance of when the controller accesses memory.
Critical weaknesses
  • The paper is poorly organized.
  • The research gap is unclear from introduction. Although the authors cite memory systems and interpretability studies, they provide limited discussion of prior work on memory's conditions of use. As a result, it is difficult to pinpoint whether the paper addresses a genuinely new problem and what its contributions are compared to existing methods.
  • The paper lacks clarity about basic terms such as scene, cue, and memory. A background section defining these central concepts would help readers understand the problem setting.
  • The discussion of related works is limited.
  • Limited test settings. The central claim is supported much more broadly within one setting than across settings. Most editing, timing, recovery, and prediction evidence comes from one K2 checkpoint on Shape×Color. The other policies support different pieces of the argument rather than reproducing the complete finding. So it is unclear how strongly the paper can generalize to robot memory overall.
Questions for authors

None

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

Related papers

Discussion (0)

Sign in to comment.