PaperlayerView exact revision on arXiv

Source-linked AI summary

EmbodiSkill: Skill-Aware Reflection for Self-Evolving Embodied Agents

Ruofei Ju, Xinrui Wang, Xin Ding, Yifan Yang, Hao Wu, Shiqi Jiang, Qianxi Zhang, Hao Wen, Xiangyu Li, Weijun Wang, Kun Li, Yunxin Liu, Haipeng Dai, Wei Wang, Ting Cao

arXiv:2605.10332v2cs.AI

TL;DR

Embodied skill self-evolution must distinguish defective guidance from execution lapses in variable environments. EmbodiSkill performs skill-aware reflection and targeted revision, achieving 93.28% ALFWorld task success with Qwen3.5-27B and improving over direct GPT-5.2 and skill-unaware evolution.

  • Problem

    Existing skill self-evolution methods often apply coarse updates, while embodied failures can arise from execution lapses rather than incorrect skill content.

    Source: §1 · p. 2, §2.2 · p. 4

  • Method

    EmbodiSkill compares each trajectory with the current skill, generates targeted reflections, and selectively updates or preserves implicated skill content.

    Source: §1 · pp. 2–3

  • Results

    93.28% ALFWorld task success with Qwen3.5-27B outperforms direct GPT-5.2 by 31.58% and skill-unaware evolution by 19.04% relative improvement.

    Source: §1 · p. 3, §4.3 · p. 13

  • Takeaways & Limitations

    Skill-aware, targeted revision helps embodied agents progressively accumulate more complete, accurate, and executable procedural knowledge.

    Source: §1 · p. 3

Abstract

from arXiv · show

Embodied agents can benefit from skills that guide object search, action execution, and state changes across diverse environments. Since embodied environments vary across layouts, object states, and other execution factors, these skills must self-evolve from trajectories generated during task execution. However, existing skill self-evolution methods are mainly developed in digital environments and often convert trajectories into coarse skill updates. Directly applying this paradigm to embodied settings is problematic, because a failed task execution may reflect not only incorrect skill content, but also an execution lapse in which the agent fails to follow valid guidance. We propose EmbodiSkill, a training-free framework for embodied skill self-evolution through skill-aware reflection and targeted revision. EmbodiSkill interprets each trajectory with respect to the current skill, uses skill-changing evidence to update the skill body, and uses execution-lapse evidence to preserve and emphasize valid guidance. Experiments on ALFWorld and EmbodiedBench show that EmbodiSkill consistently improves embodied task success. On ALFWorld, EmbodiSkill enables a frozen Qwen3.5-27B executor to reach 93.28% task success, outperforming GPT-5.2 used as a direct agent without skills by 31.58%. These results show that skill-aware self-evolution helps embodied agents accumulate reusable procedural knowledge from their own trajectories.

1. Introduction

EmbodiSkill addresses the difficulty of evolving embodied skills when failures may reflect either defective guidance or execution lapses. It uses skill-aware reflection and targeted revision to preserve valid guidance while improving skill content, achieving strong task success.

  • Embodied skills must guide recurring procedures while adapting to varied layouts, object states, visibility conditions, and feasible action sequences.The paper defines a skill as a persistent, revisable procedural specification covering prerequisites, subgoals, affordances, search, preconditions, and recovery.Source: §1 · p. 1
  • Coarse whole-skill updates are problematic because failed trajectories can reflect execution lapses rather than incorrect skill content.Such updates may add redundancy, overwrite valid guidance, or preserve incorrect prescriptions.Source: §1 · p. 2
  • EmbodiSkill examines trajectories against the current skill and classifies evidence as discovery, optimization, skill defect, or execution lapse.These targeted reflections determine what to add, optimize, correct, or preserve.Source: §1 · pp. 2–3
  • The resulting evolution loop integrates justified reflections into revised skills, which guide later executions and accumulate increasingly complete, accurate, and executable procedural knowledge.Source: §1 · p. 3
  • 93.28% task success on ALFWorld with Qwen3.5-27B outperforms direct-agent GPT-5.2 by 31.58% and G-Memory by 25.01%.EmbodiSkill also improves 19.04% relative to skill-unaware evolution.Source: §1 · p. 3

2. Related Work

Related work spans trajectory-level memory, reusable embodied skills, and self-evolving procedural knowledge. EmbodiSkill targets the remaining gap by making skill evolution selective and attribution-aware for physically grounded environments.

2.1. Memory-Based Methods for Embodied Agents

  • Memory-based methods retrieve past trajectories, summaries, or notes, but their records can remain fragmented and trajectory-specific rather than persistent procedural guidance.This motivates moving beyond retrieved memory toward skills that specify how agents should act in future executions.Source: §2.1 · p. 3

2.2. Skills and Skill Self-Evolution for Agents

  • Reusable skills and programmatic policies connect high-level instructions with executable actions, but prior methods mainly select, compose, or execute existing skills rather than revise them from trajectories.Examples include SayCan, Code as Policies, and ProgPrompt.Source: §2.2 · p. 4
  • Skill self-evolution methods extract, refine, and reuse procedural knowledge from interaction experience, including executable skill libraries and transferable trajectory-derived skills.Examples include Voyager, Trace2Skill, and EvoSkills.Source: §2.2 · p. 4
  • Embodied environments complicate coarse updates because outcomes depend on perception, spatial grounding, object states, action preconditions, and execution reliability.EmbodiSkill instead generates targeted revision signals and integrates justified changes through a skill-aware evolution spiral.Source: §2.2 · p. 4

3. Method

EmbodiSkill models embodied execution as trajectories generated under a current persistent skill. Its method repeatedly uses those trajectories for skill-aware reflection and revision.

3. Method

  • The method forms an evolution spiral: the executor uses the current skill to generate trajectories, reflections are consolidated into revisions, and the revised skill guides later execution.Source: Figure 2 · p. 4

3.1. Problem Formulation

  • An embodied trajectory records a task instruction, sequential observations and actions, and a final success signal over a finite interaction horizon.Source: §3.1 · p. 5
  • At evolution step n, the executor is guided by S(n), a persistent and revisable procedural specification for embodied task execution.The skill representation includes a prescriptive body for future task execution.Source: §3.1 · p. 5

appendix that emphasizes valid content in 𝑆(𝑛)

EmbodiSkill evolves a skill through trajectory-specific reflection, targeted body revision, and appendix updates that preserve valid guidance. A fixed executor repeatedly uses the revised skill, creating a spiral of further evidence and improvement.

3.2.1. Skill-Aware Reflection

  • Skill-aware reflection produces targeted records linking trajectory evidence and directives to Discovery, Optimization, SkillDefect, or ExecutionLapse cases.These records identify implicated existing content or propose new content; reliable revision evidence may be absent.Source: §3.2 · p. 6, §3.2.1 · p. 7
  • Discovery adds uncovered content, Optimization improves valid content, and SkillDefect corrects incorrect, incomplete, or underspecified content.ExecutionLapse instead identifies valid guidance that the executor failed to follow and creates appendix reminders.Source: §3.2.1 · p. 7

3.2.2. Skill Revision

  • Reflections are consolidated before revision by removing redundancy, merging overlaps, grouping target-specific changes, and resolving or discarding conflicting signals.This produces a consistent revision set rather than forcing uncertain changes into the skill.Source: §3.2.2 · p. 8
  • Body revision selectively edits implicated content: Discovery adds content, while Optimization and SkillDefect modify targets, leaving unrelated skill content substantively unchanged.The evolution model acts as a constrained editor, allowing limited consistency edits instead of rewriting the whole skill.Source: §3.2.2 · p. 8
  • ExecutionLapse reflections update only the appendix, which highlights valid skill content, merges duplicates, removes obsolete reminders, and incorporates new lapse evidence.They do not introduce, delete, or rewrite rules in the skill body.Source: appendix that emphasizes valid content… · p. 5, §3.2.2 · pp. 8–9

3.2.3. Skill-Aware Evolution Spiral

  • Across evolution steps, the revised skill guides later executions and generates new trajectories that expose further missing content, better strategies, defects, or execution lapses.The body becomes more complete and accurate, while the appendix makes valid content more salient during execution.Source: §3.2.3 · p. 9

4. Experiments

Experiments evaluate EmbodiSkill across embodied benchmarks with fixed executors and evolving skills learned from training trajectories. EmbodiSkill achieves stronger task success than direct agents and memory-based methods, while ablations show that targeted, skill-aware revision drives the gains.

4.1. Experimental Setup

  • The experiments use ALFWorld, EmbodiedBench-Habitat, and EmbodiedBench-Navigation, reporting task success after evolving skills on training trajectories and fixing them during held-out evaluation.Executors remain fixed during skill evolution, so gains are attributed to the evolving external skill rather than parameter updates.Source: §4.1 · pp. 9–10

4.2. Main Results

  • EmbodiSkill reaches 93.28% ALFWorld task success, outperforming GPT-5.2 direct execution by 31.58% and the strongest memory baseline by 25.01%.It also achieves the best average scores on EB-Habitat and EB-Navigation, exceeding both memory-based and direct-agent baselines.Source: §4.2 · pp. 10–12
  • EmbodiSkill performs best or ties for best on five of six ALFWorld subtasks, with Puttwo reaching 100.00% versus 52.94% for G-Memory under the same executor.The gains span task types rather than concentrating in one category.Source: §4.2 · p. 11

4.3. Ablation and Analysis

  • Skill-aware revision raises ALFWorld success from 73.13% for the static skill and 78.36% for skill-unaware evolution to 93.28%, a 19.04% relative improvement over the latter.The ablation attributes the additional gain to identifying implicated skill content and revising it in a targeted manner.Source: Table 3 · p. 12, §4.3 · p. 13
  • Across 10 revision stages, EmbodiSkill rapidly improves from 73.13% and remains in a high performance range, whereas skill-unaware evolution converges lower and fluctuates more.The stage-wise pattern indicates more stable improvement when revisions are targeted to skill content.Source: Figure 3 · p. 12, §4.3 · p. 13

5. Conclusion

EmbodiSkill enables embodied skill self-evolution through skill-aware reflection and targeted revision. Its evolution spiral uses revised skills to guide execution and subsequent trajectories to provide further improvement evidence.

  • EmbodiSkill uses skill-aware reflection to examine trajectories against the current skill and produce targeted revision signals rather than rewriting the whole skill.Source: §5 · p. 13
  • The Skill-Aware Evolution Spiral integrates revised skills into subsequent task execution, allowing new trajectories to provide further evidence for skill improvement.Source: §5 · p. 13
  • Experiments on ALFWorld and EmbodiedBench show improved task success over direct agents and memory-based baselines, while ablations confirm the importance of skill-aware revision.Source: §5 · p. 13
Loading 2605.10332v2…