acceptodds
Under review as a conference paper at ICLR 2027

World-Action-Retrospection Model for Consequence-Aligned Reuse of Experience

31 views

Abstract

Local-context world–action models (WAMs) couple visual prediction with control but leave task-critical earlier interactions outside their decision context. Memory-augmented embodied policies extend access to history, yet retaining past information does not establish whether a recalled action remains appropriate. The central challenge is to use ongoing interaction as evidence for deciding which past behavior to reuse and how it should shape action generation. We introduce WARM (World–Action–Retrospection Model), organized around event-grounded retrospection: current context and situated recollection ground inference of a candidate-independent transition requirement, while a cross-episode sensorimotor repertoire supplies state–action–effect events. Adapted repertoire actions are assessed by comparing their predicted effects with this learned forecast; consequence-gated source transport then jointly regulates proposal-informed stochastic initialization and candidate-derived conditioning, turning recalled behavior into a revisable prior for flow-based action generation. WARM achieves 98.7% success on LIBERO and 83.9% on RMBench, including 87.3% on tasks requiring multiple past observations. Component comparisons suggest complementary benefits from recollection and repertoire; complete-event distraction tests show less degradation with explicit consequence comparison. These findings support using current evidence to govern both the selection of past behavior and its influence on action generation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robot learning, world-action models, manipulation, event-grounded retrospection, consequence-gated source transport, flow matching, imitation learning

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_rkWYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how robots use the current task history and past experience to generate actions. The method uses within-episode memory to infer what needs to change, retrieves actions from previous episodes, and assesses whether their predicted effects fit the current situation. A shared gate controls how much the retrieved experience influences both the conditioning and the flow initialization. Experiments on LIBERO, RMBench, and real-world robot tasks, together with retrieval distraction tests and component comparisons, show benefits from selectively reusing past experience.

Critical strengths
  1. The problem is meaningful, and the motivation is clear. An action that worked before may not be appropriate now, especially when the robot needs to track task progress or decide whether to repeat an interaction. The method goes beyond retrieving similar experience and checks whether it fits the current need, which is useful for robot control.

  2. The approach allows flexible reuse of experience. Retrieved actions serve as references that the model can adapt to the current scene, and their influence can be reduced or removed when they are not useful. This lets the robot benefit from past experience while remaining responsive to the current situation.

Critical weaknesses

The overall advantage and generalization ability are still unclear. RMBench uses task-specific models and configurations, and retrieval includes task identity. The overall gain over the strongest baseline is small, with clearer improvements concentrated in some tasks requiring multiple past observations. The results therefore provide stronger support for experience reuse within known tasks than for a unified policy that handles new tasks.

The evaluation/demos in this paper are also limited in scope, as all the experiments focus on tabletop manipulation.

Questions for authors
  1. How much does the system depend on the quality and diversity of past experience? How closely must the stored examples match the current task to be helpful? For example, if a plate is in a different position from where it was in the retrieved demonstration, how does this affect action adaptation and task success?

  2. When the experience bank lacks relevant skills, rather than just containing some misleading candidates, can the gate reliably suppress the retrieved experience? Does performance remain at least comparable to a policy without retrieval in this setting?

Reviewer_Hh4EOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how to improve WAM's performance with augmentation from the history of execution as well as a bank of example demonstrations. The paper proposes WARM, which identifies important experiences during the current execution, retrieves demonstrations from the bank to generate future predicted consequences, and guides the final action generation. The paper compares WARM with baselines on LIBERO, RMBench, and real-robot tasks to demonstrate the usefulness of the retrieved information.

Critical strengths
  1. The paper studies an important and realistic scenario for WAM: how to identify the most important memory from both the ongoing execution and the training demonstrations to aid the current action generation. WARM proposes a comprehensive pipeline to incoporate those information.
Critical weaknesses
  1. The improvement compared to MemoryWAM is not very significant. MemoryWAM seems to be the most relevant baseline to WARM; however, it is only compared under the RMBench evaluation in the main experiments. Based on the experiment, the improvement of WARM could not be fully justified based on the reported numbers.

  2. The computational costs seem to be missing. How is WARM's cost compared to baselines such as FastWAM and MemoryWAM?

  3. The proposed method seems to be an overly engineered pipeline to me, with many components: episode memory, demonstration retrieval, action adaptation, consequence prediction, candidate ranking, gating, etc. The training loss also contains many components.

  4. Currently, the retrieval part may not be heavily tested. For example, for the real-robot experiments, there are very limited demonstrations, so retrieval could be relatively trivial, while in practice the method may need to retrieve from a more general repertoire with potentially a diverse set of demos.

  5. The training objective of the candidate regularizer needs more justification. Currently, it supervises different candidates with the same effects and it does not directly teach the different consequences of those alternatives.

Questions for authors

Please address the weaknesses part.

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

This paper proposes WARM, a world-action model that uses past experience to guide robotic manipulation. It maintains two types of memory: situated recollection from the current episode and a repertoire of actions and their observed effects from previous episodes. The model estimates the transition required by the current task, evaluates retrieved actions based on their predicted consequences, and uses a shared gate to control their influence on flow-based action generation. WARM reports 98.7% success on LIBERO and 83.9% on RMBench, with additional experiments on physical robots.

Critical strengths
  1. The paper tackles a practically relevant problem: actions that work well in one setting can turn out unsuitable when reused under different circumstances.
  2. Its consequence‑driven action selection offers a solid extension to prior retrieval‑based policies including ReCAP and informed‑source flow approaches.
  3. The distinction between in‑episode recollection and cross‑episode experience is sensibly motivated.
  4. Ablation studies back up the complementary roles played by recollection, the repertoire, and consequence‑aware selection, with the effect especially visible on RMBench.
  5. Beyond standard benchmark numbers, the distractor perturbation tests and real‑robot evaluations supply valuable supporting evidence.
Critical weaknesses
  1. Gains against top‑performing baselines remain modest. WARM hits 98.7% on LIBERO compared to WLA‑0’s 98.6%, and 83.9% on RMBench versus MemoryWAM’s 83.0%. It is left unproven whether these gaps carry statistical significance.
  2. Comparisons against existing experience‑conditioned policies lack proper experimental controls. Approaches like ReCAP and informed‑source flow also leverage retrieved actions or learned action priors. While the text discusses conceptual differences to WARM, there are no head‑to‑head runs using a shared backbone and equal training budgets.
  3. Candidate‑level outcomes cast doubt on how well consequence prediction actually works. As shown in Table 4, predicted effects lift applicable‑action selection merely from 17.6% up to 19.0%. Even with extra diagnostics given in the appendix, it stays ambiguous how this fairly weak selection component contributes to the overall end‑to‑end policy performance.
  4. Robustness testing has a narrow scope. Misleading‑experience perturbation experiments are only run over two tasks: Put Back Block and Press Button. We cannot tell if the reported benefits would generalize across a wider set of manipulation scenarios.
  5. The overall pipeline comes with substantial complexity. It intertwines recollection, retrieval, action adaptation, effect prediction, ranking and gated generation. Inference‑time overhead has not been well quantified, which makes it hard to judge real‑world practical trade‑offs.
  6. Real‑robot validation is still restricted. Four tasks with 20 trials apiece offer partial proof of transfer capability, yet these results are not enough to confirm robustness under larger shifts of objects and working environments.
Questions for authors
  1. The overall improvements over WLA‑0 on LIBERO and MemoryWAM on RMBench are relatively small. Could the authors provide uncertainty estimates or matched‑protocol comparisons to clarify whether these gains are meaningful? Why does WARM underperform MemoryWAM on the M(1) task group?
  2. Table 4 reports only 19.0% applicable action selection using predicted effects. The appendix further reports 41.7% candidate‑pool coverage and 27.3% all‑query selection accuracy. How often does the action generator recover from an unsuitable retrieved proposal, and how much does this recovery contribute to the final success rate?
  3. The misleading‑experience experiments cover only two RMBench tasks. Does consequence‑based selection remain effective on other tasks, particularly when the proportion of irrelevant retrieved actions increases beyond 50%?
  4. WARM introduces several additional modules compared with its Fast‑WAM backbone. What is the inference‑time overhead of retrieval, effect prediction, and gating? Is the improvement over simpler experience‑conditioned action generation sufficient to justify this additional cost?
1 comment

Thanks for the additional experiments. The new comparisons and analysis help clear up several of my concerns, especially about the contribution of the proposed method and its inference cost. I still think the experiments could cover more tasks, and a more direct comparison with related methods would be useful. But overall, the response is convincing enough for me to keep my Weak Accept rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.