acceptodds
Withdrawn from ICLR 2027

Recover, Don't Just Continue: Learning from Feedback for Physical Failure Understanding and Correction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal large language models are increasingly seen as the cognitive foundation for embodied intelligence, yet planning from instructions alone cannot make them reliable embodied brains. During physical execution, actions deviate from intended effects, and what can be realized is jointly constrained by the embodiment, controller, and environment. An embodied brain must therefore understand why execution deviates, recognize what the current execution system can reliably realize, and revise decisions using interaction experience. These three needs correspond to causal failure localization, capability-constrained adaptation, and experience-driven revision, making them complementary and non-substitutable. However, current embodied foundation models, despite advances in spatio-temporal reasoning, physical-world modeling, and high-level planning, largely lack these capabilities. To address this gap, we introduce Embodied-Corrector. Controlled failure–recovery data connects actions, observable consequences, failure causes, and executable recoveries, providing physical causal supervision. Evaluation-first SFT with policy-optimization internalization transfers verified evaluations into the model's own reasoning, so it assesses outcomes and failure causes before inferring recoveries a frozen executor can realize, enabling execution-capability adaptation. A CoT template requires the model to review execution history, evaluate failure causes, and plan recovery, so each execution informs the next decision without parameter updates, realizing in-context learning from interaction. This internalizes diagnostic and corrective experience into the model's reasoning. Experiments on embodied failure reasoning and planning benchmarks show stronger failure-cause analysis and corrective planning; with a frozen VLA, composite task success improves by 6.1%, and with fixed GPT-5.5 execution, embodied manipulation and navigation tasks improve by 13.9% macro-average.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Embodied AI, Embodied Reasoning, Execution Correction, Long-Horizon Robot Manipulation

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_j6rxOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Embodied-Corrector, a framework for understanding physical execution failures and generating corrective plans in embodied robotic tasks. The authors construct controlled failure–recovery data from successful trajectories and introduce an evaluation-first training strategy followed by Noise-Aware Entropy-Gated Policy Optimization (NEPO). The proposed method is evaluated on failure reasoning and planning benchmarks, as well as robotic execution tasks with fixed code-generation and VLA backends, demonstrating improvements in failure diagnosis, corrective planning, and task success.

Critical strengths

The paper addresses an important problem in embodied control by emphasizing physical failure understanding and corrective planning rather than simply detecting failures or continuing execution. The controlled failure–recovery data construction provides structured supervision connecting observed execution outcomes, failure causes, and corrective actions. The evaluation-first training strategy and NEPO objective are well motivated by the mismatch between verified evaluations during training and model-generated evaluations during inference. The experiments cover multiple benchmarks and execution settings, including frozen VLA and code-generation backends, providing evidence of the effectiveness of the proposed approach.

Critical weaknesses

The training data are explicitly constructed from recoverable failures. While this is a reasonable assumption, the paper would benefit from a clearer discussion of the intended scope and practical value of the proposed method. In particular, it would be helpful to clarify under what realistic robotic scenarios this recoverability assumption holds and how the proposed approach provides meaningful benefits in such settings. Furthermore, in practical deployment, it is important to consider when the system should appropriately terminate execution rather than repeatedly attempting recovery. This is not a criticism of the method's inability to handle irreversible failures, but rather a request to better clarify its applicability, operational boundaries, and practical significance.

In embodied control, the consequences of an incorrect action may not become apparent immediately but instead manifest several steps later. While the proposed dataset considers missing prerequisites, it remains unclear whether the model can trace an observed failure back to an earlier incorrect decision when there is a substantial temporal gap between the cause and its manifestation. I believe this delayed causal attribution is an important aspect of physical failure understanding that deserves more direct investigation and evaluation.

While the paper provides comparisons across foundation models and post-training objectives and includes a self-correction baseline, direct comparisons with existing failure-recovery approaches discussed in the Related Work section are largely missing. Evaluating representative methods under comparable execution settings would help clarify the advantages of the proposed approach over existing correction mechanisms and better establish its technical contribution.

Although multiple execution trials are conducted, the paper does not report independent training runs across multiple random seeds or sufficient uncertainty estimates. Some evaluation conditions also involve relatively few trials. These limitations make it difficult to assess the robustness and statistical significance of the reported improvements, especially when gains over strong baselines are relatively small.

Questions for authors

Please see the weaknesses.

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

This paper proposes Embodied-Corrector to identify the causes of robot execution failures and generate recovery plans. The authors construct failure data through controlled perturbations, first train outcome judgment and cause analysis, then recovery planning, and further optimize the model using NEPO. The experiments cover failure diagnosis, task planning, and closed-loop simulation tasks with frozen execution models. The results show improvements on some composite tasks, but success rates may decrease on atomic tasks.

Critical strengths
  1. The data construction approach is reasonable. Changing only one execution factor at a time helps identify failure causes. Splitting training and test trajectories before generating perturbations also reduces the risk of data leakage.
  2. The experiments include closed-loop simulation execution, not just text-based scoring. Keeping the execution models fixed helps assess the correction module’s contribution. The authors also report performance degradation on atomic tasks, making the results more complete.
Critical weaknesses
  1. Real-robot experiments are missing, and practical applicability remains unverified. The paper emphasizes physical failure understanding and adaptation to execution capabilities, but all execution experiments are conducted in simulation.
  2. Evaluation details and key comparisons are insufficient. The paper does not adequately explain the diagnostic test set size, the Cause/Fine scoring rules, or NEPO’s reward calculation and key parameters. Training uses both failure data and general planning data, while execution combines task decomposition, cause diagnosis, and feedback from execution history. However, comparisons separating these effects are missing.
Questions for authors
  1. Could the authors provide real-robot results to show whether the method works beyond controlled simulation?
  2. With the same amount of data and computational budget, what results are achieved by using only general planning data, and by performing task decomposition without providing failure causes?
Reviewer_iWCrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an embodied error-correction model trained on controlled failure and recovery data to assess execution outcomes, identify failure causes, and plan corrective actions. By combining supervised fine-tuning with entropy-gated policy optimization, it improves performance on complex tasks with a fixed executor. However, the experiments are conducted entirely in simulation, performance declines on simpler tasks, and real-world deployment remains unvalidated.

Critical strengths

A program finishing execution does not necessarily mean that the physical task has been completed. For example, a gripper may close without grasping an object, leaving subsequent transport and placement actions without a necessary precondition. This paper addresses the mismatch between a plan’s assumptions and the actual state of the environment. It requires the system to verify execution outcomes and restore unmet preconditions before proceeding. This is a concrete problem that directly affects the success rate of long-horizon tasks.

The approach starts from successful trajectories and modifies a single factor at a subtask boundary, such as grasp height, gripper timing, or the target object instance, before observing the execution outcome. Compared with asking a language model to infer failure causes after the fact, this approach makes the intervention explicit while preserving the same successful execution prefix. This facilitates establishing links between interventions, observed failures, and recovery actions.

With the models, visual inputs, and action interfaces held constant, the study compares three feedback conditions: no additional feedback, feedback indicating only the failure outcome, and feedback explaining the failure cause. Providing causal feedback increases the success rates of the three models from 22% to 28%, from 20% to 27%, and from 20% to 24%, respectively. These results directly motivate the proposed approach: feedback should contain information that helps the system determine how to modify its actions.

Critical weaknesses

All experiments are conducted in simulation, leaving real-world effectiveness unverified. Appendix B.2 explicitly states that all quantitative execution experiments are performed in simulated environments. The conclusion also acknowledges that observation noise and state uncertainty in real-world settings remain unresolved. The proposed method relies on accurately determining states such as whether an object has been grasped or a door has been opened sufficiently. In practice, these judgments may be affected by occlusion, camera viewpoints, and sensor errors. The current results therefore demonstrate error-correction effectiveness in simulation but do not establish its effectiveness on real robots.

The evaluation criteria for open-ended responses are insufficiently specified. Table 1 reports Cause, Sub, and Fine as percentages without adequately explaining how these language responses are judged to be correct. A single failure may admit multiple valid recovery strategies, yet it is unclear whether responses must match a reference answer or are evaluated using rules, human judgment, or another model. Fine also covers actions, parameters, execution order, and verification steps, but the treatment of partially correct responses is not clearly described.

Mixing general planning data with error-correction data makes the source of some improvements difficult to isolate. Appendix B.1 states that supervised fine-tuning uses both controlled error-correction data and general planning data from Embodied-Reasoner. Although Table 2 shows substantial improvements on EB-ALFRED, the paper does not compare training on general planning data alone, error-correction data alone, and their combination. Consequently, these results cannot establish how much of the improvement in planning comes from learning from failures versus additional general planning supervision.

The computational and time costs of the correction pipeline are not reported. The system invokes the error-correction model and then passes its output to a fixed GPT model to generate the subsequent workflow. Although the paper emphasizes that the correction model produces all four fields in a single response, it does not report the additional latency, number of model calls, token consumption, or total task completion time. For robotic systems, higher success rates may come at the cost of longer waiting times and more replanning attempts. Given the average improvement of 3.7 percentage points over GPT-based self-correction, the absence of cost measurements makes it difficult to assess whether the gains justify the additional overhead in practical use.

Questions for authors

see weaknesses

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

Related papers

Discussion (0)

Sign in to comment.