acceptodds
Withdrawn from ICLR 2027

EMHO: EMbodied Agent Harness Optimization via Experience Traces

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Improving embodied agents often focuses on optimizing the underlying model through training, while the surrounding agent harness that controls planning, context, and tool use is typically engineered. We ask whether this harness can instead improve itself directly from experience traces under sparse environmental feedback. We propose EMbodied Agent Harness Optimization (EMHO), a self-evolving framework that keeps the embodied model frozen and iteratively revises its harness by analyzing execution trajectories and prior harness history. EMHO optimizes beyond skills or recovery prompts, modifying how the agent monitors progress, uses vision tools, grounds observations, and responds to failures. To support multiple subtasks with a single harness, we introduce EMHO-Merge, which addresses trade-offs in jointly optimizing a single shared harness across subtasks by using episode-level gains and losses to guide evidence-supported refinement of when and how revised behaviors are applied. We evaluate EMHO on EmbodiedBench across navigation and manipulation tasks, and EMHO consistently improves task success for both Qwen 9B and 27B models. Qualitative analysis shows that EMHO goes beyond recovering from failures and unproductive actions to reshape how the embodied agent interprets and interacts with its environment.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Embodied Agent, Harness Optimization, Self Evolving

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_3LZcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes EMbodied Agent Harness Optimization (EMHO), a framework that automatically optimizes the embodied agent harness using execution traces and historical harness modifications while keeping the underlying model frozen. EMHO iteratively analyzes task execution experiences, modifies harness components such as planning, memory, tool usage, and failure recovery, and retains modifications that improve task success rates. The paper further introduces EMHO-Merge, which uses episode-level gains and losses to refine a shared harness across multiple subtasks and mitigate performance regressions. Experiments on EmbodiedBench across navigation and manipulation tasks demonstrate improvements over the initial harness for both Qwen 9B and 27B models, with additional comparisons against one-shot and skill-only optimization baselines.

Critical strengths

The paper addresses an important and meaningful research problem: automatically optimizing the entire embodied agent harness rather than relying on model training or optimizing individual components. This research direction is well-motivated and has significant potential for improving embodied agents.

Critical weaknesses
  1. EMHO repeatedly uses the same Search Set throughout the optimization process, making it prone to overfitting and lacking effective mechanisms to ensure generalization.

  2. EMHO generates only one candidate from the current best Harness in each iteration and accepts the modification only if the candidate achieves a strictly higher success rate on the Search Set. This search strategy limits exploration and may prematurely discard promising optimization directions.

  3. EMHO relies on continuously accumulated experience traces and harness modification histories for optimization. However, large amounts of redundant, outdated, or conflicting information may interfere with the optimizer's decision-making and increase computational overhead.

  4. EMHO heavily relies on an LLM to perform a series of tasks, including trajectory analysis, failure diagnosis, modification planning, and Harness code editing, making the process susceptible to hallucinations and error propagation. Moreover, Harness editing operations are primarily restricted through prompt-based soft constraints on the editable scope, which are unreliable.

  5. Although the paper emphasizes that EMHO can optimize the entire Harness, it does not systematically investigate or evaluate the evolution of workflow structures, such as component execution order and stage nesting. Therefore, its claim of full-Harness optimization is insufficiently substantiated.

  6. Table 1 reports the best Search Set and Held-out Set results separately across optimization iterations, while Appendix B further acknowledges that these results may come from different iterations. Therefore, the reported best Held-out performance does not necessarily correspond to the final harness selected based on Search Set performance and cannot accurately reflect the generalization performance of EMHO.

Questions for authors
  1. Table 3 does not report the Held-out results of EMHO (Combined), making it difficult to assess whether EMHO-Merge improves generalization over standard shared-harness optimization. Moreover, EMHO-Merge performs worse than EMHO (Separate) on both the manipulation Search and Held-out sets. Could the authors provide the missing results and clarify the actual benefits of EMHO-Merge?

  2. Table 2 compares full-harness optimization with skill-only optimization exclusively on the Search Set, without reporting Held-out performance. Moreover, full-harness optimization does not consistently outperform skill-only optimization across subtasks. Could the authors provide Held-out comparisons to substantiate the claimed advantages of full-harness optimization?

  3. EMHO performs 10 iterations of experience collection, harness revision, and evaluation, whereas the one-shot baselines are allowed only a single revision. This gives EMHO substantially more opportunities to exploit execution feedback and improve performance. Does EMHO still outperform these baselines under matched optimization and task-execution budgets?

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

The paper proposes EMHO, which improves a frozen embodied VLM by iteratively editing its harness. The same model uses execution traces and edit history to propose patches, retaining those that improve search-set success. EMHO-Merge builds a shared harness across subtasks, using gained and lost successes to guide a targeted repair. On EmbodiedBench navigation and manipulation with search/held-out splits, EMHO outperforms the initial harness, skill-only optimization, and one-shot rewrites by frontier models. Discovered edits improve recovery, progress monitoring, tool selection, and visual grounding.

Critical strengths
  1. Timely question with a broad edit space. The paper moves harness optimization from coding agents to embodied agents with sparse feedback. Unlike prior work that edits only skills or recovery prompts, EMHO edits executable runtime code, and the skill-only comparison suggests this larger space helps.
  2. Concrete analysis of what the optimizer discovers. Section 5 links specific code and prompt edits to measurable behavior changes on held-out trajectories.
Critical weaknesses
  1. Held-out numbers are not a valid generalization estimate. Table 1 reports the maximum held-out score over all iterations, which may come from a different harness than the one selected on the search set. Appendix B itself calls these "retrospective summaries" rather than generalization estimates.
  2. Methodological novelty is limited. The optimization loop (read traces and edit history, propose a patch, keep it if search success improves) closely follows Meta-Harness and Self-Harness. The new elements are mainly the embodied setting and the gain/loss repair in EMHO-Merge, whose benefit is not supported on held-out data (see 3).
  3. EMHO-Merge's benefit is shown only on the search set. On held-out, Merge equals Separate in navigation (75.6 vs. 75.6) and is worse in manipulation (33.8 vs. 40.7). EMHO (Combined) has no held-out results.
  4. Cost is not reported. There are no numbers for optimizer tokens, environment episodes or wall-clock time, which matters when comparing against a single one-shot rewrite.
Questions for authors
  1. What is the held-out success of the harness selected only by search-set score (the final accepted parent) for every row of Tables 1 to 3?
  2. Can you report mean and std over at least 3 independent optimization runs, and state the number of episodes per split for each subtask?
  3. What is technically different from Meta-Harness and Self-Harness beyond the application domain? Can these methods be applied directly to EmbodiedBench as baselines, and how does EMHO compare to them?
  4. Can you report held-out results for EMHO (Combined) in Table 3?
  5. What is the total cost of one EMHO run, and how many of the 10 proposed patches are accepted or fail to run?
Reviewer_eCwwOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

EMHO improves embodied agent performance by improving the harness around the frozen vision-language model. There are two loops. Inner loop collect harness experiences, and the outer loop tries to optimize based on the set of harness experiences and picks the best one. Authors experiment on EmbodiedBench navigation and manipulation tasks using Qwen models as backbone.

Critical strengths
  1. Problem definition is clear, and the motivation is easy to understand
  2. Section 5 and Appendix D provide some good insights into the behaviour changes of a small open model optimizing its own harness.
Critical weaknesses
  1. The paper lacks significant novelty and looks purely engineered.
  2. Authors don’t talk about the limitations of their work and where it fails.
  3. Evaluation scope is very narrow. Experiments were done only on EmbodiedBench and that too on just navigation and manipulation.
  4. Methodology section is pretty vague. The optimizer in eq 3 is somewhat treated as a black box. How history traces become evidence is unclear.
  5. Authors say: EMHO entries report the best score attained on each split over the optimization iterations. The results could be cherry-picked. How many seeds the experiments were run for table 1,2, and 3 and what are the standard deviation or confidence interval of these scores.
  6. No comparison with the closest prior methods, i.e., Meta-Harness, Self-Harness, AutoSaddler, EmbodiSkill, ASPIRE, and Zetta
Questions for authors
  1. Can the authors clearly justify the novelty? What does the framework actually contribute other than prompt design and accept-if-better selection?
  2. Weakness 5
  3. Weakness 6

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

Related papers

Discussion (0)

Sign in to comment.