acceptodds
Under review as a conference paper at ICLR 2027

EmbodiExp: Compositional Experience Transfer for Embodied Agents

Abstract

Experience can help embodied agents turn high-level plans into grounded actions, but transferring it across tasks remains challenging: previously effective strategies may be inapplicable to the current state or fail to produce their intended outcomes. We present EmbodiExp, a framework for constructing transferable experiences and composing them online during embodied task execution. Our key observation is that tasks often share subgoals whose completion can be assessed independently of overall task success. We therefore introduce an outcome-grounded subgoal experience that specifies a subgoal, applicability conditions, an adaptable policy, and a verifiable outcome. During execution, applicable experiences guide action generation, while verified outcomes can establish the conditions required by subsequent experiences, enabling online composition. Subgoal-level outcome attribution further enables failure utilization: failed trajectories can still contribute verified local successes and condition-specific guidance on actions to avoid. Experiments on multiple embodied benchmarks under a fold-based protocol with frozen experience banks demonstrate the effectiveness of EmbodiExp.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied Agents, Experience Transfer, Agent Memory

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_is1ZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes EmbodiExp to improve embodied task execution by reusing experience from past interactions. The method abstracts local attempts within trajectories into a four-field representation consisting of a subgoal, applicability conditions, an execution policy, and an intended outcome. It extracts effective behaviors and condition-specific avoidance guidance from both successful and failed trajectories. During execution, the system retrieves relevant experiences, checks their applicability, generates actions suited to the current environment, and verifies local outcomes using observable evidence. These outcomes inform task-state updates, enabling the composition of experiences from different sources. The experience bank is constructed before evaluation and subsequently used during evaluation. Experiments on EB-Habitat, EB-ALFRED, and VIGIL demonstrate notable overall improvements.

Critical strengths
  1. The experience representation has clear execution semantics. The method provides operational guidance, checks applicability conditions, and verifies the outcomes of actual execution. Retaining locally successful experience from failed trajectories is also practically valuable.
  2. The experimental results are reasonably convincing. Compared with NoExp, success rates increase from 50.0% to 68.0% on EB-Habitat and from 54.7% to 68.3% on EB-ALFRED, while world-state completion on VIGIL improves from 31.6% to 40.8%. The method also outperforms the evaluated memory systems and embodied experience-based methods.
  3. The paper includes ablations on experience granularity, applicability checking, outcome verification, and the use of failed trajectories. It also reports cross-model experience transfer and both positive and negative transfer. The frozen-bank, three-fold evaluation protocol clearly separates experience construction from evaluation on held-out tasks.
Critical weaknesses
  1. The incremental contribution over the closest related work remains insufficiently clear. Although the paper compares against WorldMind and EmbodiSkill, it lacks a direct comparison or sufficiently detailed analysis of the methodological differences from skill-based approaches such as MSCE, which incorporate supporting evidence, applicability boundaries, and verification rules.
  2. The evidence for generalization remains limited in scope. Each of the three environments uses its own experience bank, so results across multiple benchmarks do not establish that a single bank can transfer across environments or action interfaces. Moreover, the EB-Habitat and EB-ALFRED splits are not fully scene-disjoint or task-template-disjoint.
  3. Comparisons regarding model configurations and compute budgets remain incomplete. The paper specifies the execution backbones but does not clearly identify the models used for experience construction, applicability checking, state maintenance, and outcome verification. In particular, the configurations of the online auxiliary modules in the 8B execution experiments require clarification. The current results therefore do not fully disentangle the contribution of experience reuse from that of additional inference-time computation.
  4. Negative transfer and the reliability of outcome verification are insufficiently analyzed. On EB-Habitat, 62 episodes that fail under NoExp become successful, but eight previously successful episodes regress to failures. The paper reports this phenomenon without providing specific explanations for the regression cases.
Questions for authors
  1. Please clarify how EmbodiExp differs from MSCE, WorldMind, and EmbodiSkill in experience representation, applicability conditions, outcome verification, and composition mechanisms. Could you provide a direct comparison with the closest approaches, or compare different experience representations under a shared execution and verification pipeline?
  2. Could you separately report performance on new combinations of previously encountered subgoals and on unseen subgoals or operation types, together with experience hit rates, applicability-check pass rates, and the proportion of execution performed without experience guidance? Please distinguish cross-task transfer within an environment from cross-environment transfer. If the latter has not been evaluated, the corresponding claims should be qualified.
  3. Please identify the specific model used for each module, particularly whether all online auxiliary modules in the 8B experiments also use an 8B model or introduce a stronger model. Could you add a NoExp comparison under the same token or model-call budget and separately report online execution and offline construction costs?
3 comments

We thank the reviewer for the constructive feedback. We have revised the manuscript and uploaded the updated version.

Weakness 1 & Question 1: Comparison with MSCE, WorldMind and EmbodiSkill

Direct Comparison. We have added MSCE to the baselines in Table 1, with its adaptation to our frozen-bank protocol detailed in Appendix B.6. EmbodiExp outperforms MSCE, WorldMind, and EmbodiSkill on EB-Habitat and EB-ALFRED.

MethodEB-Habitat (%)EB-ALFRED (%)
NoExp50.054.7
WorldMind51.755.7
MSCE51.054.7
EmbodiSkill52.351.0
EmbodiExp68.068.3

Methodological Differences. We have expanded the Experience and Skill Learning paragraph in Related Work to clarify the methodological distinctions. MSCE promotes evidence-supported policies into callable skills with applicability boundaries and verification rules. WorldMind learns causal rules from prediction errors and planning heuristics from successful trajectories. EmbodiSkill selectively revises persistent procedural skills by distinguishing skill defects from execution lapses.

EmbodiExp constructs subgoal experiences through local outcome attribution, whereas MSCE uses reflection-weighted propagation of terminal feedback to govern skill promotion. During reuse, newly verified outcomes can establish the conditions for applying experiences from other tasks, enabling online composition.

Weakness 2 & Question 2: Scope and Analysis of Generalization

Scope. We evaluate compositional experience transfer across tasks within each environment, using a separate frozen experience bank for each environment. Transfer of a single bank across environments or action interfaces has not been evaluated. We have clarified this scope in Appendix D.

Cross-Scene and Cross-Task Transfer. Our original three-fold split balances task categories across folds. We have added two separate EB-ALFRED evaluations to assess transfer to unseen scenes and task templates. The scene-disjoint evaluation uses three scene-grouped folds, rebuilding each experience bank after excluding all test-scene trajectories. The task-template-disjoint evaluation uses 300 held-out tasks whose templates are excluded from experience construction. In each setting, NoExp, A-Mem (the strongest baseline in the original EB-ALFRED evaluation) and EmbodiExp are evaluated on the same test tasks. The results are added to Table 9.

MethodScene-disjoint SR (%)Task-template-disjoint SR (%)
NoExp54.330.7
A-Mem59.336.0
EmbodiExp69.062.3

EmbodiExp outperforms A-Mem by 9.7 percentage points on unseen scenes and 26.3 percentage points on unseen task templates, supporting both cross-scene and cross-task experience transfer within EB-ALFRED.

Subgoal Generalization and Experience Use. We conducted a separate supplementary evaluation on 300 EB-Habitat episodes, matching tasks, folds, and initial observations between NoExp and EmbodiExp. This supplementary run achieved overall SRs of 50.0% and 70.0% for NoExp and EmbodiExp, respectively. Within each fold, tasks are grouped by whether their individual subgoals and complete goal configurations appear in the source tasks referenced by its experience bank. Grounded subgoals distinguish the goal condition, object category, and target location.

Task groupEpisodesNoExp SR (%)EmbodiExp SR (%)Hit-rate bounds (%)Applicability pass (%)Actions without guidance (%)
Seen subgoals, new combinations6552.372.357.4–66.296.540.1
Unseen grounded subgoals14446.566.756.0–65.596.741.7
Seen subgoals, seen combinations9153.873.636.4–43.496.256.5

Hit-rate bounds reflect incomplete relevance logging: selector-confirmed relevant retrievals (lower) and nonempty retrievals (upper), each divided by all queries. Applicability and unguided-action rates use checked candidates and submitted environment actions, respectively.

Gains persist on new goal configurations: improvements are comparable for new and seen combinations (+20.0/+19.8 pp), while tasks containing unseen grounded subgoals improve from 46.5% to 66.7%.

Unseen Operation Types. Unseen grounded subgoals can involve familiar operations applied to new object categories or target locations. The analysis above supports transfer to new grounded goals and goal combinations, but does not establish transfer to previously unseen operation types.

Weakness 3 & Question 3: Model Configurations and Compute

Model Configurations. The main experiments use Qwen3.6-35B-A3B on EmbodiedBench and Qwen3.6-27B on VIGIL for experience construction and all online LLM modules. In Figure 4A, the constructor (Qwen3.6-35B-A3B or GPT-5.4) operates exclusively offline, while the selected executor handles all online operations, including applicability checking, Task State maintenance, and outcome verification. All online modules in the 8B experiments use Qwen3-VL-8B. These configurations are specified in Appendix B.

Contribution of Experience Reuse. We added an empty-bank control on EB-Habitat that retains the same executor, action limits, and Task State, outcome-verification, and replanning prompts as EmbodiExp. Revised Table 8 reports success rates and average online costs per episode.

MethodSR (%) ↑Env. stepsModel callsTokens (k)
NoExp50.015.967.3437.84
EmbodiExp (empty bank)50.716.0221.7874.56
EmbodiExp (Ours)68.013.6717.4477.22

The empty-bank control achieves 50.7% SR, close to NoExp’s 50.0%, despite using 96.6% of EmbodiExp’s tokens and making more model calls. Adding the experience bank improves SR by 17.3 percentage points with comparable token usage and fewer calls, supporting the contribution of experience reuse beyond additional online computation.

Offline Costs. On EB-Habitat, source-trajectory collection and experience-bank construction together require 2,187 model calls and 10.87M tokens in total across the three folds. These costs include collecting the 300 shared source trajectories and constructing the fold-specific experience banks, each using 200 source trajectories. Local MiniLM indexing adds no generative-model calls or tokens. These offline costs are reported separately from online execution costs in Appendix C.2.

Weakness 4: Negative Transfer and Outcome Verification

Negative Transfer. We analyzed the eight episodes that succeed under NoExp but fail under EmbodiExp and added the observed failure modes to Section 4.3.

Failure modeCases
Budget exhaustion during search or repeated attempts3
Manipulation of non-target objects3
Violation of global task constraints1
Execution deviating from the selected policy1

Outcome Verification. We reviewed 1,121 recorded verification decisions against the corresponding action and environment-feedback logs, checking whether the task’s required target satisfied the subgoal at the time of each decision.

Verifier decisionChecksReview result
Subgoal completed359357 supported by evidence; 2 incorrect (0.56%)
Completion not confirmed759No missed completions identified in the recorded evidence
Unknown3No action had been executed
Reviewer_GU2BOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes EmbodiExp, a framework that stores past interactions of embodied agents as subgoal-level experiences and reuses them on new tasks. Each experience contains a subgoal, the conditions under which it applies, execution guidance and an outcome that can be verified. An LLM constructor segments source trajectories into subgoal attempts, attributes local outcomes and admits entries from both successful and failed trajectories. At test time, the agent maintains a Task State, retrieves candidate experiences for the current subgoal, applies the highest ranked one whose conditions hold and credits the outcome only after verification. Verified outcomes can enable later experiences, which the authors call online composition. Under a three-fold protocol with frozen banks, EmbodiExp improves success from 50.0 to 68.0 on EB-Habitat, from 54.7 to 68.3 on EB-ALFRED and from 31.6 to 40.8 on VIGIL. It also outperforms six memory and experience baselines.

Critical strengths
  1. The evaluation is conducted in a careful manner. The experience bank is built only from episodes outside the test split and stays fixed during evaluation. The splits are made without looking at results, every method uses the same splits and the agent has no access to the benchmark's success signals. The results are consistent across splits. On EB-Habitat, EmbodiExp solves 62 episodes that NoExp fails and fails only 8 that NoExp solves, which makes the overall improvement convincing.
  2. Attaching to each piece of guidance the conditions under which it applies and an outcome that can be checked is a simple and practical idea. The ablations show that both the applicability check and the use of failed trajectories make a clear difference, which is useful for others building embodied agents that learn from past experience.
Critical weaknesses
  1. It is unclear how much of the improvement comes from the experience itself. EmbodiExp also adds a Task State, per-step outcome checks and a replanning prompt. It uses about 2.4 times as many model calls as NoExp, but the paper never tests this pipeline with an empty bank. Experiences from different trajectories are chained together in only 11.7% of EB-Habitat episodes, so the composition mechanism explains little of the 18 point improvement.
  2. On EmbodiedBench, the setting is closer to reusing experience from similar tasks than to real transfer. The bank comes from the same model's runs on the other 200 episodes, which may share scenes and task templates with the test episodes. On VIGIL, where test scenes are unseen, the margin over the best baseline drops to 4.4%. All results come from a single run and some baselines fall below NoExp without explanation.
Questions for authors
  1. Could you report NoExp with the same Task State, verification and replanning prompts but with an empty bank? This would show how much of the improvement comes from the experience itself.
  2. The Subgoal-Level Experience and Similarity-Based Composition rows in Table 3 are identical. Could you clarify which components each of these rows includes?
  3. Could you show how success changes with how closely each test subgoal matches its nearest bank entry? Results on EB-ALFRED with unseen scenes would also be helpful.
  4. Could you explain how the baselines were adapted to a fixed experience bank?
5 comments

We thank the reviewer for the constructive feedback. We have revised the manuscript and uploaded the updated version.

Weakness 1 & Question 1: Contribution of Experience and Composition

We have added an empty-bank control that retains the same Task State, outcome verification, and replanning prompts as EmbodiExp, using the same executor and action limits. The results are added to the revised Table 8.

MethodSR (%) ↑Env. stepsModel callsTokens (k)
NoExp50.015.967.3437.84
EmbodiExp (empty bank)50.716.0221.7874.56
EmbodiExp (Ours)68.013.6717.4477.22

The empty-bank control achieves 50.7% SR, close to NoExp’s 50.0%, despite using 96.6% of EmbodiExp’s tokens and making more model calls (21.78 vs. 17.44). Under the same execution pipeline and similar online token usage, adding the experience bank improves SR by 17.3 percentage points, supporting the contribution of experience beyond the additional execution scaffold.

Role of Composition. The 18.0-point overall improvement reflects the full framework, which supports both individual experience transfer and composition. The 11.7% statistic specifically counts episodes in which verified outcomes establish the conditions for subsequent experiences drawn from different source trajectories. Such composition occurs more frequently in Long-Horizon tasks (28.0%; Figure 4C). The ablations in Table 3 further show that incorporating applicability checks and outcome verification improves SR from 58.3% to 68.0% over similarity-only subgoal reuse, supporting the value of state-grounded experience use.

Weakness 2 & Question 3: Experience Similarity and Transfer to Unseen Scenes and Task Templates

Our original three-fold EmbodiedBench split balances task categories across folds but does not enforce scene or task-template separation. We have clarified this scope and added the following experiments and analyses.

Experience Similarity. Our focus is compositional experience transfer: tasks with different overall goals can share subgoals, allowing subgoal-level experiences from prior tasks to be reused and combined under new task requirements.

To examine local matching, we analyzed 848 executed subgoal attempts with specific descriptions and recorded outcome checks from a 300-episode EB-Habitat run. For each attempt, we computed the maximum MiniLM cosine similarity between its pre-selection subgoal description and the subgoal descriptions in the corresponding frozen bank, then divided the attempts into three approximately equal-sized groups.

Nearest-entry similaritySubgoal completion rate (%)
Low28.27
Medium29.68
High32.62

Completion increases modestly from 28.27% to 32.62%, showing a positive association between local similarity and completion in this sample.

Transfer to Unseen Scenes and Task Templates. We have added two separate EB-ALFRED evaluations to assess transfer to unseen scenes and task templates. The scene-disjoint evaluation uses three scene-grouped folds, rebuilding each experience bank after excluding all test-scene trajectories. The task-template-disjoint evaluation uses 300 held-out tasks whose templates are excluded from experience construction. In each setting, NoExp, A-Mem (the strongest baseline in the original EB-ALFRED evaluation), and EmbodiExp are evaluated on the same test tasks. The results are added to Table 9.

MethodScene-disjoint SR (%)Task-template-disjoint SR (%)
NoExp54.330.7
A-Mem59.336.0
EmbodiExp69.062.3

EmbodiExp outperforms A-Mem by 9.7 percentage points on unseen scenes and 26.3 percentage points on unseen task templates, supporting both cross-scene and cross-task experience transfer within EB-ALFRED.

VIGIL Results. We acknowledge that the margin over the strongest baseline is smaller on VIGIL. EmbodiExp achieves 40.8%, exceeding NoExp (31.6%) and the strongest baseline, ReMe (36.4%), by 9.2 and 4.4 percentage points, respectively. One possible explanation is that VIGIL requires agents to independently determine whether their actions have succeeded, a challenge that experience alone may not fully address. The additional EB-ALFRED evaluations above provide complementary evidence of transfer to unseen scenes and task templates.

Variability across Runs. We repeated the EB-Habitat evaluation with three seeds. EmbodiExp outperforms NoExp in every run by 14.0-22.3 percentage points, with mean SRs of 68.33% and 50.78%, respectively. These results, reported in revised Table 10, support the consistency of the improvement across the evaluated runs.

MethodSeed 101Seed 102Seed 103Mean ± SD
NoExp51.33%54.00%47.00%50.78 ± 3.53%
EmbodiExp65.33%70.33%69.33%68.33 ± 2.65%

Baselines below NoExp. As reported in Table 1, EmbodiSkill underperforms NoExp on EB-ALFRED (51.0% vs. 54.7%). This may reflect negative transfer when prior guidance does not match the current task conditions. The result highlights that experience reuse does not uniformly improve execution, and we have added this discussion to the EmbodiedBench results analysis.

Question 2: Identical Rows in Table 3

Subgoal-Level Experience and Similarity-Based Composition report the same reference configuration: subgoal-level experiences with similarity-based retrieval, without online applicability checks or outcome verification. The granularity study varies the experience unit among trajectory, action, and subgoal levels under this setting. The composition study fixes subgoal-level experiences and varies the two online checks. These rows therefore provide a common reference for the two ablation studies. We have clarified this in the ablation descriptions and the caption of Table 3.

Question 4: Fixed-Bank Adaptation of Baselines

We have added a Baseline Adaptation subsection in Implementation Details. Each method constructs its memory offline from source trajectories, retaining its memory representation. The resulting memory remains frozen during held-out evaluation, with no persistent updates from test episodes.

A-Mem, Mem0, and LangMem construct memory notes, inferred memories, and structured procedures, respectively, from successful and failed trajectories. ReMe distills experiences from successful trajectories only. At episode start, A-Mem retrieves up to three task-relevant entries, while Mem0, LangMem, and ReMe each retrieve up to five, following their respective retrieval procedures.

WorldMind constructs goal and process experiences offline and retrieves up to three goal and five process experiences per planner call. Its test-time reflections are retained only within the current episode and do not modify the frozen bank. EmbodiSkill constructs its manual through offline reflection and uses the frozen manual to guide planning during evaluation.

Reviewer_9dXMOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents EmbodiedExp, a framework for using existing trajectories, successful and unsuccessful, to perform better on tasks even when the task has changed. The specific mechanism is to store outcome-grounded subgoal experiences, a datatype parametrized by a subgoal, applicability conditions, adaptable policy, and verifiable outcome. Chaining subgoals, even from other tasks and contexts, should lead to valid trajectories. The paper goes through the framework and experimental methodology. It then shows results on EmbodiedBench and VIGIL. It uses held-out folds to show transfer capability. Use of EmbodiedExp leads to outperformance of skill/memory-related baselines; ablations of each component underperform as well.

Critical strengths

Strength 1: paper idea is intuitive and sensible. I think this format lends itself especially well to large multimodal models. Strength 2: idea itself is communicated clearly. Strength 3: I have novelty concerns (see below) but the experimentation done here, because it uses current models, is a valuable contribution. The results framework is extensive and fairly well-selected - aside from a major oversight I discuss below, the baselines and ablations are well-chosen and the results are convincing.

Critical weaknesses

Weakness 1: this idea does not seem novel - experience units seem to be more or less equivalent to PDDL actions. The process seems to be inferring PDDL actions from existing trajectories and applying them to new tasks, all using foundation models. This has been tried before [1]. The experience unit format in particular, which is presented as not one but two contributions (I think it should be one), appears to be essentially a PDDL action and therefore not at all novel.

Weakness 2: related to weakness 1, pddl-style subgoal-related methods should at least be in both the related work and the baselines.

Weakness 3: the writing is at times unclear and at other times overly repetitive. After reading the abstract I was confused - I wasn't quite sure what the paper was talking about as it immediately entered jargon without explaining itself. But then when I read the intro and methods I felt as though the same ideas were being repeated at me.

[1] "One Demo Is All It Takes: Planning Domain Derivation with LLMs from A Single Demonstration". Jinbang Huang, Yixin Xiao, Zhanguang Zhang, Mark Coates, Jianye Hao, Yingxue Zhang. ICLR 2026.

Questions for authors
  • How does this work distinguish itself from other subgoal-style works, especially those using LLMs?
  • When constructing an experience, how do you differentiate between what to abstract away and what to take as general? For example in Figure 1, it makes sense that the specific object being put in the drawer is not in the experience, but I don't fully understand why "drawer" is in the experience. That could make sense but the general principle (even one the LLM follows and the authors observe) isn't communicated to the reader. Is it because the drawer is part of the environment? Does that principle generalize? Is it something the LLM came up with rather than the authors?
4 comments

We thank the reviewer for the constructive feedback. We have revised the manuscript and uploaded the updated version.

Weakness 1: Novelty and Relationship to PDDL

We agree that the intuition of relating behavior to applicability conditions and outcomes is shared with symbolic planning. Our contributions to experience learning and transfer include an outcome-grounded experience representation and mechanisms for constructing and composing experiences across tasks.

PDDLLM induces predicates and action models from demonstrations and physical simulation for symbolic task-and-motion planning. EmbodiExp extracts subgoal experience from local action-outcome evidence, retaining verified successes within failed tasks and deriving condition-specific avoidance guidance from explicit action failures. During reuse, applicability is checked against current evidence, and newly verified outcomes can establish the conditions for applying experiences from other tasks.

We have clarified these contributions as follows:

  • We present EmbodiExp, a framework for constructing outcome-grounded subgoal experiences from past task executions and transferring them to new tasks.
  • We develop online experience composition across different trajectories, where verified local outcomes from one experience can establish the applicability conditions of another.
  • We propose subgoal-level outcome attribution for failure utilization, retaining locally successful behavior and deriving condition-specific avoidance guidance from explicit failure evidence.

Weakness 2: Missing PDDL-Style Related Work and Baselines

We have expanded the Related Work section to cover PDDL-based planning and domain-learning methods, including PDDLLM and UniDomain [1], and clarified their relationship to our approach.

We have added UniDomain baseline, as we could not locate a runnable public implementation of PDDLLM. We made substantial efforts to reproduce UniDomain as faithfully as possible using its public implementation. The adaptation uses Qwen3.6-35B-A3B and fold-isolated source demonstrations. Results are included in Table 1, with adaptation details in Appendix B.6.

MethodEB-Habitat (%)EB-ALFRED (%)
NoExp50.054.7
UniDomain17.016.0
EmbodiExp68.068.3

Failure Analysis. On EB-ALFRED, the learned domains lacked an explicit heating effect, and the planner returned no plan for 25 of the 28 heat-and-place episodes. On EB-Habitat, 160 episodes produced a plan but ended in task failure, with no simulator-confirmed successful manipulation recorded in those episodes. These observations characterize learned-domain coverage and execution outcomes under the evaluated adaptation.

[1] "Pretraining a Unified PDDL Domain from Real-World Demonstrations for Generalizable Robot Task Planning". Haoming Ye, Yunxiao Xiao, Cewu Lu, Panpan Cai. NeurIPS 2025.

Weakness 3: Writing Clarity and Repetition

We have rewritten the abstract as below to make our motivation and method easier to follow. We have also consolidated repeated explanations across the Introduction and Method sections.

Experience from past tasks can help embodied agents perform new tasks, but it may not remain applicable or effective when conditions change. Tasks with different overall goals often share subgoals, providing a basis for cross-task experience transfer. We present EmbodiExp, a framework for constructing outcome-grounded subgoal experiences and composing them in new tasks. Each experience describes how to pursue a subgoal, when the guidance applies, and what outcome must be verified for completion. Verified outcomes from applying one experience can satisfy the conditions for applying another, allowing experiences from different past tasks to be composed online. By examining the actions and observed outcomes at the subgoal level, EmbodiExp retains experience from successful subgoals and uses explicit action failures to identify actions to avoid under specific conditions. Experiments on EmbodiedBench and VIGIL assess cross-task transfer with experiences kept fixed throughout held-out evaluation. EmbodiExp achieves success rates of 68.0% on EB-Habitat and 68.3% on EB-ALFRED, compared with 50.0% and 54.7% for the agent without stored experience.

Question 1: Distinction from Other LLM-Based Subgoal Methods

EmbodiExp uses subgoals as the units for extracting and transferring experience across tasks. Beyond identifying intermediate goals or retrieving subgoal plans as examples, it specifies what experience is supported by local action-outcome evidence, when that guidance applies, and what outcome must be verified. Verified outcomes from one experience can establish the conditions for applying another, enabling online composition across different trajectories.

We have expanded the Multimodal Embodied Agents and Experience and Skill Learning paragraphs in Related Work to clarify this focus and our contributions in experience representation, construction, and compositional reuse.

Question 2: Experience Abstraction

We specify the abstraction principle in our experience-construction prompts (Appendix E.1): remove incidental source-specific details while retaining properties and task constraints needed for applicability, execution, and the intended outcome. The rule is specified by the authors, and the LLM applies it to individual local attempts. It applies equally to manipulated objects and environmental entities.

For example, if a placement experience applies to different types of open containers, “drawer” can be generalized to “container”, with openness retained as an applicability condition. Any drawer-specific properties needed for applicability, execution, or the intended outcome should remain explicit.

As a motivating example of compositional experience transfer, Figure 1 uses abbreviated subgoal labels instead of complete experience specifications. To make the illustrated abstraction clear, we revised the labels to “open container”, “put object in container”, and “pick up object”, while retaining the concrete task descriptions.

Reviewer_9dXMOpenReview ↗

Improving my score

Thanks for the thorough response and follow-up results. This does prove the value of the method over similar structures and I see the difference. Improving to a 3.

Dear Reviewer 9dXM,

We sincerely appreciate your time, effort, and thoughtful feedback. Your insightful suggestions have helped us improve our manuscript.

We are grateful to you for updating your assessment and raising your rating. We are encouraged by your recognition of the method's value and appreciate your careful consideration of our responses and additional results.

Best regards,

Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Question and Response Summary

We appreciate the AC's time in reviewing our paper, responses, and discussions.

For the AC's consideration:

  1. 9dXM raised their rating from 2 (Weak rejection) to 3 (Weak acceptance) following our response and additional results, recognizing the method's value and its distinction from related approaches.
  2. GU2B stated that their questions and concerns had been mainly addressed and maintained their positive score (3).
  3. is1Z initially rated the paper 3 (Weak acceptance). We responded to all four major concerns with clarifications and additional evidence.

Strength Summary

Reviewers recognized the intuitive and sensible idea (9dXM), clear execution semantics and value of retaining local successes from failed trajectories (is1Z), and convincing empirical results (9dXM, GU2B, is1Z). GU2B and is1Z also acknowledged the careful frozen-bank evaluation.

Main Concern Summary

Novelty and Prior Work

  • Q: Distinction from PDDL and existing subgoal/skill methods? (9dXM, is1Z)
    • [Clarification] We clarified local outcome attribution, failure utilization, and online composition: verified outcomes can establish the conditions for applying experiences from other trajectories. In follow-up, 9dXM acknowledged the method's value over similar structures and recognized the distinction.
    • [Additional material] We expanded Related Work and added UniDomain and MSCE comparisons (Table 1; Appendix B.6). We made substantial efforts to reproduce UniDomain using its public implementation, as we could not locate a runnable public implementation of PDDLLM.

Experience Reuse and Composition

  • Q: Do gains come from experience or additional online computation? (GU2B, is1Z)
    • [Additional material] An empty-bank control with the same execution pipeline achieves 50.7% SR versus 68.0% with experience, despite comparable token usage and more model calls, supporting the contribution of stored experience (Table 8).
  • Q: What does the 11.7% composition rate imply? (GU2B)
    • [Clarification] This counts cross-source composition, which reaches 28.0% in Long-Horizon tasks. Overall gains also include individual experience reuse. Applicability checking and outcome verification improve SR from 58.3% to 68.0% over similarity-only reuse (Table 3).

Generalization and Robustness

  • Q: Transfer to unseen scenes and task templates? (GU2B, is1Z)
    • [Additional material] Separate EB-ALFRED evaluations yield 69.0% scene-disjoint and 62.3% task-template-disjoint SR, exceeding A-Mem by 9.7 and 26.3 percentage points (Table 9).
  • Q: Experience similarity, new subgoal combinations, and unseen subgoals? (GU2B, is1Z)
    • [Additional material] Similarity analysis over 848 attempts found a modest positive association with completion. A supplementary 300-episode analysis reported gains over NoExp on new combinations (52.3% to 72.3%) and unseen grounded subgoals (46.5% to 66.7%), alongside experience-use statistics.
    • [Clarification] Single-bank transfer across environments has not been evaluated; results on unseen grounded subgoals do not establish transfer to unseen operation types.
  • Q: Are improvements consistent across runs? (GU2B)
    • [Additional material] Three-seed EB-Habitat evaluation shows gains in every run, with mean SRs of 68.33% versus 50.78% for NoExp (Table 10).

Model Configurations, Costs, and Reliability

  • Q: Models for auxiliary modules and offline/online costs? (is1Z)
    • [Clarification] We specified module configurations; all online modules in the 8B experiments use Qwen3-VL-8B (Appendix B).
    • [Additional material] We separately reported offline collection/construction and online execution costs (Appendix C.2; Table 8).
  • Q: Negative transfer and outcome-verification reliability? (is1Z)
    • [Additional material] We analyzed all eight regression cases and reviewed 1,121 verification decisions against execution logs. Of 359 positive completion decisions, 357 were supported by evidence and two were incorrect (0.56% of positive decisions).

Other Question Summary

  • Q: Unclear or repetitive writing? (9dXM)
    • [Clarification] We rewrote the abstract and consolidated repeated explanations.
  • Q: Experience abstraction principles? (9dXM)
    • [Clarification] We clarified the author-defined rule: remove incidental details while retaining properties needed for applicability, execution, and outcomes. We also revised Figure 1.
  • Q: Frozen-bank baseline adaptation and identical Table 3 rows? (GU2B)
    • [Clarification] We documented baseline adaptations and explained that the two rows share a reference configuration.

Notes:

  • [Clarification]: Explanations and presentation revisions.
  • [Additional material]: New experiments, analyses, and comparisons during rebuttal.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.