acceptodds
Under review as a conference paper at ICLR 2027

Reflect After Failure: Brain-Inspired Self-Evolving Agents for Embodied Spatial Intelligence

Abstract

Vision-Language Models (VLMs) have made substantial progress in spatial reasoning, and self-evolving agents further extend these capabilities to embodied environments by learning from interactions. However, current self-evolution remains particularly challenging in embodied spatial tasks, complex and multi-step decisions from partial visual evidence making it difficult to determine why an execution fails. The failure may stem from misperception, insufficient exploration, or incorrect planning, yet execution feedback alone rarely reveals which component is responsible. As a result, simply accumulating interaction experience can reinforce ineffective strategies. To address this, we propose HiMeco, a brain-inspired self-evolving agent for embodied spatial intelligence, inspired by metacognitive monitoring and hippocampal memory consolidation. Through Hippocampal-Metacognitive Reflection, HiMeco diagnoses failures in perception–exploration–planning (PEP), then compares execution attempts to derive reusable strategies and guidance for avoiding recurring errors. It also abstracts interaction programs into parameterized skills, using evidence contracts and execution-based validation to determine their retains and reuse. These strategies and skills guide subsequent interactions, with memory updated only between rounds and model parameters kept frozen. HiMeco achieves state-of-the-art performance on ESI-Bench, and it also improves accuracy on EmbSpatial-Bench by 3.94 percentage points without evolution.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Brain-Inspired, Hippocampal-Metacognitive Reflection, Self-Evolving agent, Embodied Spatial Intelligence

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_UF4ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HiMeco, a self-evolving embodied agent that combines perception–exploration–planning (PEP) failure diagnosis with two forms of external memory: a strategy bank and a library of executable skills. After each interaction round, HiMeco uses execution traces, task outcomes, and diagnostic feedback to identify failure causes, consolidate reusable strategies, abstract parameterized skills, and validate them through later executions. The VLM parameters remain frozen throughout evolution. Experiments on ESI-Bench and EmbSpatial-Bench report substantial improvements over active VLM baselines.

Critical strengths

1.The paper addresses an important issue in self-evolving embodied agents: failed interaction alone does not reveal whether the underlying problem comes from perception, exploration, or planning. Explicitly separating these failure sources is a reasonable and potentially useful direction.

2.The framework is well structured. Strategy consolidation, skill abstraction, evidence contracts, and execution-based skill validation form a coherent evolution pipeline, rather than simply storing successful trajectories in memory.

3.The reported performance gains are substantial. On the selected ESI-Bench episodes, HiMeco improves GPT-5.6 from 56.89% to 69.19% and Gemini-3-Flash from 54.74% to 62.51%. The paper also provides component ablations and detailed implementation information.

Critical weaknesses

1.The main evaluation repeatedly adapts on the same ESI-Bench episodes, making it difficult to separate genuine generalization from instance-specific adaptation. The cross-episode transfer evidence is mainly qualitative.

2.The PEP diagnosis relies on privileged retrospective information, including simulator metadata and benchmark correctness. This limits the applicability of the proposed evolution mechanism in realistic settings where such supervision is unavailable.

3.The SOTA comparison is not fully controlled because different methods are evaluated on different episode subsets or sample counts. This weakens the strength of the benchmark-level claim.

4.The ablation study covers only a subset of tasks and does not fully disentangle the contributions of diagnosis, strategy memory, skill evolution, and evidence validation.

5.The “brain-inspired” framing is stronger than the technical evidence. The implemented mechanisms largely resemble structured reflection, external memory, and skill consolidation, while the biological analogy is mainly motivational.

Questions for authors

Clarification requests

  1. Can the authors clarify whether the main ESI-Bench results measure generalization or repeated-instance adaptation, given that the same episodes are revisited across evolution rounds?

  2. Which parts of the PEP diagnosis require privileged simulator metadata or benchmark correctness, and how much of the method would remain applicable without this information?

  3. How should the SOTA comparison be interpreted when some methods are evaluated on different episode subsets or sample counts?

  4. Do the component-ablation conclusions hold across all nine ESI-Bench tasks, or only the four tasks reported in Table 2?

Suggestions for future work

None. My current concerns are primarily about clarification and evaluation validity rather than missing future extensions.

Reviewer_FffsOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HiMeco, a self-evolving framework for embodied spatial reasoning. The authors categorize failures into perception, exploration, and planning errors, and construct a diagnostic module using execution traces and privileged simulator information to provide structured feedback. Based on this feedback, the agent iteratively refines its strategies and executable skills without updating model parameters. Experiments on ESI-Bench show improved performance through repeated interactions, with additional evaluations on EmbSpatial-Bench and limited cross-episode generalization examples.

Critical strengths
  1. Clear failure categorization. The paper identifies three failure modes in embodied spatial reasoning: perception, exploration, and planning. This provides a structured perspective for diagnosing agent failures and guiding improvements.

  2. Clear framework design and presentation (minor strength). The overall pipeline is well organized and clearly presented. However, I consider this a minor strength rather than a substantive research contribution, as the framework itself does not introduce a fundamentally new research question.

Critical weaknesses
  1. Limited generalization (major concern). The main experiments perform self-evolution and evaluation on the same 208 episodes. The reported 12.30% improvement primarily reflects in-domain adaptation, with substantial risk of benchmark-specific overfitting. Only a few qualitative transfer examples are provided, without systematic evaluation on unseen episodes or environments.

  2. Limited novelty in the research question. Failure diagnosis and reflection-guided agent improvement have already been explored in prior work [1, 2, 3]. While the paper introduces a perception–exploration–planning categorization, it does not establish a fundamentally new research question or sufficiently distinct insight.

  3. Limited methodological novelty. The framework mainly combines existing techniques, including heuristic failure diagnosis, reflection, and skill refinement. The resulting design is relatively complex without a correspondingly clear or novel underlying principle, making its contribution appear primarily incremental.

[1] Duan et al., AHA: A Vision-Language-Model for Detecting and Reasoning over Failures in Robotic Manipulation, ICLR 2025.

[2] Ju et al., EmbodiSkill: Skill-Aware Reflection for Self-Evolving Embodied Agents, 2026.

[3] Zhang et al., Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence, 2026.

Questions for authors
  1. How well does HiMeco generalize to completely unseen datasets or out-of-domain spatial reasoning tasks (e.g., BEHAVIOR-1K [4] and MolmoSpaces-Bench [5])? Can the authors provide quantitative evidence that the learned strategies and skills transfer beyond ESI-Bench, rather than primarily adapting to benchmark-specific patterns?

  2. Could the authors more clearly articulate the key conceptual and methodological novelty of HiMeco compared with existing work on failure diagnosis, reflection, and skill evolution [1–3]? In particular, what fundamentally new insight or capability does the proposed framework introduce?

[1] Duan et al., AHA: A Vision-Language-Model for Detecting and Reasoning over Failures in Robotic Manipulation, ICLR 2025.

[2] Ju et al., EmbodiSkill: Skill-Aware Reflection for Self-Evolving Embodied Agents, 2026.

[3] Zhang et al., Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence, 2026.

[4] Li et al., BEHAVIOR-1K: A Benchmark for Embodied AI with 1,000 Everyday Activities and Realistic Simulation. CoRL 2022.

[5] Kim et al., MolmoSpaces: A Large-Scale Open Ecosystem for Robot Navigation and Manipulation. arXiv:2602.11337, 2026.

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

This paper proposes HiMeco, a brain-inspired self-evolving agent framework for embodied spatial reasoning. The key motivation is that existing self-evolving agents may accumulate ineffective experiences because execution failures are difficult to attribute to specific causes in embodied environments. To address this issue, HiMeco introduces a Perception–Exploration–Planning (PEP) diagnostic module that analyzes execution trajectories, task outcomes, and privileged simulator metadata to identify potential failure sources. Based on these diagnostics, the framework compares multiple execution attempts and consolidates experiences into two forms of memory: a strategy bank containing reusable guidance and negative constraints, and a skill library containing parameterized executable programs. Evidence contracts and execution-based validation are further employed to regulate skill retention and reuse. The framework enables memory-based self-evolution without updating the underlying VLM parameters.

Critical strengths

S1. Clear and meaningful motivation. The paper identifies an important challenge in self-evolving embodied agents: execution feedback alone may be insufficient to determine whether failures originate from perception, exploration, or planning. The observation that incorrect failure attribution can lead to ineffective experience accumulation provides a reasonable motivation for the proposed framework. S2. A coherent framework integrating failure diagnosis and memory evolution. HiMeco integrates structured failure diagnosis, comparative reflection, strategy consolidation, and executable skill abstraction into a unified framework. In particular, maintaining both strategy-level and skill-level memories allows the agent to reuse high-level behavioral guidance and concrete interaction routines. The use of evidence contracts and execution-based validation is also a reasonable design choice for reducing the retention of unreliable experiences. S3. Promising empirical improvements on embodied spatial reasoning. The experimental results demonstrate substantial improvements over the corresponding active baselines on ESI-Bench, with gains of 12.30 and 7.77 percentage points for GPT-5.6 and Gemini-3-Flash, respectively. The ablation study further indicates that the strategy bank, skill library, and PEP-guided reflection contribute to the overall performance. The reported reductions in execution time on most evaluated tasks are also encouraging.

Critical weaknesses

W1. Limited methodological novelty beyond existing reflection- and memory-based self-evolving agents. The core components of HiMeco, including LLM-based reflection, experience memory, executable skill abstraction, and execution-based skill validation, have been extensively explored in previous research. The main distinction appears to be the integration of structured PEP diagnostics into the memory evolution process. However, the PEP diagnostic module itself is largely implemented through manually designed heuristics, fixed thresholds, and an LLM auditor, rather than a substantially new learning or reasoning mechanism. The paper does not sufficiently clarify the conceptual and technical differences between HiMeco and closely related approaches. Consequently, it remains unclear whether the proposed framework represents a significant methodological advance or primarily an integration of existing techniques. W2. The reliability of PEP failure attribution is insufficiently validated. Accurate failure attribution is the central motivation of this paper, yet the PEP diagnostic module primarily relies on heuristic indicators rather than direct verification of the underlying failure causes. For example, exploration failures are assessed using reference-viewpoint proximity and pose diversity, which do not necessarily indicate whether the agent has acquired sufficient task-relevant evidence. Similarly, perception diagnosis uses answer correctness and predefined rules, making it difficult to distinguish genuine perception errors from downstream reasoning errors. A more convincing evaluation would include independently annotated failure categories or controlled failure cases to assess diagnostic accuracy and its relationship to subsequent performance improvements. W3. Insufficiently controlled comparisons with existing self-evolving methods. The main quantitative comparisons primarily involve passive and active VLM baselines, while comparisons with closely related self-evolving agents are limited. Furthermore, HiMeco benefits from multiple execution attempts and memory updates across rounds. It is therefore important to determine whether the observed improvements originate from the proposed diagnostic mechanism or simply from additional interaction experience, repeated sampling, and memory accumulation. The current experiments do not fully disentangle these factors through comparisons under matched interaction and computational budgets. W4. Dependence on privileged simulator information limits applicability. The PEP diagnostic module relies on privileged simulator metadata, including reference viewpoints and simulator-derived information, which is unavailable to the online planner. While this design is reasonable for post-execution analysis in simulation, it raises concerns about whether the proposed self-evolution mechanism can operate in real-world or less instrumented embodied environments, where such metadata may not be available. The paper does not provide a dedicated evaluation of the impact of removing privileged simulator information. As a result, the extent to which HiMeco depends on privileged supervision remains unclear.

Questions for authors

Q1. What performance can HiMeco achieve when PEP diagnosis relies only on execution traces and task outcomes? Q2. How does HiMeco compare with existing self-evolving agents under identical settings?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.