Source-linked AI summary
Self-Evolving Embodied Agents via Skill-Harness Evolution
Peidong Wang, Zhiming Ma, Ying Chang, Xufang Luo, Xiaocui Yang, Shi Feng, Yuqing Yang, Dongsheng Li
TL;DR
Parameter-updating methods adapt embodied agents but require weights, demonstrations or rewards, and additional optimization. SHAPER instead keeps planner and executor weights frozen while evolving reusable skills and a context-code harness through target-environment rollouts, improving performance across VLABench and ESI-Bench settings.
Problem
Adapting embodied agents commonly requires model-weight access, task-specific demonstrations or rewards, and additional optimization, motivating train-free adaptation of frozen agents.
Method
SHAPER evolves textual procedural skills and a context-code harness through target-environment rollouts while keeping the planner and executor parameters fixed.
Results
Across VLABench and ESI-Bench, SHAPER improves the same frozen planner across different low-level action interfaces and is reported as a competitive alternative to fine-tuning and sampling-heavy baselines.
Takeaways & Limitations
Skill-and-harness evolution provides a lightweight route for adapting frozen embodied agents when parameter updates are expensive, unavailable, or undesirable.
Takeaways & Limitations
The planner and executor remain frozen, so adaptation is restricted to optimizing the reusable skill and context-code harness.
Abstract
from arXiv · showhide
Embodied agents are increasingly built as systems around foundation models, where performance depends not only on model weights but also on the skills, context, action interfaces, and execution harness surrounding the model. While supervised fine-tuning and reinforcement learning can adapt agents to new environments, they require additional data, rewards, and training runs; meanwhile, many train-free code-centric approaches rely on programmable robot APIs that may be unavailable in fixed-interface settings. We propose SHAPER, a self-evolving framework for train-free embodied adaptation that keeps model parameters frozen and improves the non-parametric agent system by evolving reusable skills and a context-code harness through target-environment rollouts. In SHAPER, the same frozen model can serve as both planner and optimizer, refining its external skills and context-code harness without parameter updates. We evaluate SHAPER on VLABench and ESI-Bench, covering embodied agents with different low-level action interfaces, and compare against pure execution, supervised fine-tuning, and test-time-scaling baselines such as verifier-free selection and voting. Our results suggest that skill-and-harness optimization is a practical route to self-evolving embodied agents when model training is expensive, unavailable, or undesirable.
1 Introduction
SHAPER addresses train-free adaptation of frozen embodied agents by evolving reusable skills and a context-code harness from target-environment rollouts. It evaluates this approach across embodied settings with different action interfaces against parameter-updating and inference-time baselines.
- SFT and RL can improve embodied performance but require model weights, task-specific demonstrations or rewards, and additional optimization.
- Train-free robotics approaches often improve agents by generating or repairing executable programs that call robot-specific APIs.
- SHAPER instead evolves reusable skills and a context-code harness from target-environment trajectories and outcomes while keeping planner and executor weights frozen.Skills provide procedural guidance, while the harness organizes observations, prior actions, execution outcomes, and feedback for the planner.
- The same frozen VLM serves as the upper-level planner during interaction and as an artifact optimizer during evolution, using distinct prompts and inputs.
- SHAPER is evaluated on VLABench and ESI-Bench with different low-level action interfaces against pure execution, same-data SFT, and test-time-scaling baselines.
- The paper frames skill-and-harness optimization as a competitive alternative to fine-tuning and sampling-heavy baselines for adapting embodied agents without parameter updates.
2 Related Work
Related work adapts embodied agents through parameter updates, external skills, harnesses, or train-free interfaces. SHAPER extends this line by jointly evolving skills and context-code harnesses around frozen models.
- Embodied foundation-model approaches broaden generalist robot control and embodied reasoning, while reinforcement learning further expands their capabilities.
- These parameter-updating approaches depend on additional data, rewards or corrections, and model updates, unlike SHAPER’s external skill and harness artifacts.
- SkillOpt evolves a persistent natural-language skill while keeping the execution harness fixed, and EmbodiSkill evolves skills while distinguishing defective content from execution lapses.
- Train-free embodied systems use affordances, executable robot APIs, spatial value maps, or reusable skills, but code- and API-centric methods depend on platform-specific interfaces and monitoring components.
3 Method
SHAPER factorizes a frozen embodied agent into a planner, executor, reusable skill, and context-code harness, then optimizes only the external artifacts. It uses rollout diagnosis, staged evolution, sandboxed validation, and top-K selection to improve candidate artifacts.
- Agent factorization: SHAPER conditions a frozen VLM planner and executor with a reusable textual skill and context-code harness.The executor remains fixed, while the skill provides procedural guidance and the harness constructs planner context.
- Agent factorization: Train-free adaptation freezes planner and executor parameters while optimizing the external pair c = (s, h).The candidate consists of the reusable skill and harness rather than neural weights.
- Rollout evaluation: SHAPER evaluates candidate artifacts through target-environment rollouts containing executed actions, observations, and a task-level reward or success signal.Expected rollout performance is estimated on finite training and validation sets.
- Rollout-guided diagnosis: A hierarchical textual diagnosis compares each response and action with observations immediately before and after execution, then aggregates critiques, metadata, context, and outcomes.The resulting textual gradient exposes patterns such as repetition, instruction drift, missing progress tracking, harmful context, and parsing or API errors.
- Two-stage evolution: The frozen model separately acts as optimizer, proposing candidates from the current artifact, textual gradient, and optimization history.Evolution first updates the skill with the seed harness fixed, then updates the harness with the selected skill fixed.
- Validation and selection: Generated harnesses undergo sandboxed validation before rollout, and valid candidates compete with incumbents through top-K selection.Invalid candidates are rejected, while the final output is the highest-scoring candidate encountered across rounds.
4 Experiments
SHAPER is evaluated as a frozen-agent adaptation method across VLABench and ESI-Bench, using held-out splits and different execution interfaces. Results show gains from evolving skills and harnesses, with effects varying across shifts, categories, and artifact configurations.
- Experimental setup: SHAPER is evaluated with frozen planner and executor parameters across VLABench and ESI-Bench, whose agents use different low-level action interfaces.VLABench delegates textual subgoals to a frozen VLA actor, whereas ESI-Bench uses a fixed benchmark action API.
- Experimental setup: VLABench evaluation uses four 200-episode held-out splits that vary target categories and task form, including unseen-target and cross-task-form shifts.The splits are disjoint from the 15 training and 24 validation episodes used for artifact evolution.
- VLABench results: 34.50% success is achieved by full skill-harness evolution on VLABench, exceeding the 28.25% Seed Agent and same-data SFT baseline.The Seed Agent also exceeds direct VLA execution by 5.00 points and SFT by 4.25 points without changing either model.
- VLABench results: Full SHAPER gains are larger under distribution shift, reaching +6.0 points on unseen targets, +10.0 on an unseen task form, and +6.5 when both are unseen.It achieves the best result on C1, C2, and C4, and ties Skill Evolution at 50.0% on C3.
- Artifact ablations: Skill-only, harness-only, and full evolution score 33.50%, 30.50%, and 34.50%, versus 28.25% for the seed, with non-additive artifact effects.Adding the evolved harness after skill evolution changes the overall score by +1.00 point, while split-wise effects range from 0.0 to +2.5 points.
- Qualitative analysis: Controlled cases attribute improvements to progress-aware skill changes and evidence-preserving harnesses that retain task-critical visual information across later steps.The evolved skill changes commands after stalled retrieval, while the evolved harness preserves reflected targets and deterministic focus crops for matching.
- ESI-Bench results: On ESI-Bench, full evolution reaches 49.8% micro and 42.9% macro accuracy, exceeding the Seed Agent by 17.3 and 11.7 points.The macro score is numerically above the published GPT-5 PS reference, but the comparison is external because evaluation sets differ.
- ESI-Bench results: Category-level improvements are strongest for Specular Reflection, Perceptual Grounding, and Spatial Relations, while Enumerative Perception and Action Sequencing decrease.Action Sequencing contains only five questions, so category-level differences should be interpreted with sample counts.
5 Conclusion
SHAPER improves embodied agents by evolving a textual skill and context-code harness while keeping planner and executor parameters fixed. Across VLABench and ESI-Bench, the frozen 27B planner improves across different action interfaces, though cross-embodiment transfer and real-robot validation remain future work.
- SHAPER evolves a textual skill and context-code harness while keeping planner and executor parameters fixed.
- The approach uses rollout-derived textual feedback in a two-stage optimization to produce reusable procedural guidance and trajectory context rather than per-episode action searches.
- The same frozen 27B upper-level planner improves with both a VLA executor and an active-perception action interface.
- Improvements include held-out target categories and task forms, while cross-embodiment transfer and real-robot validation remain future work.
A Additional Artifact and Optimization Details
The appendix documents the artifacts and prompt templates used in the experiments, clarifying that skills and harnesses are the two optimization targets. It distinguishes persistent textual instructions from executable context-construction code.
- The appendix reproduces the artifacts and prompt templates used in the experiments.
- A skill is the planner’s persistent textual instruction.
- A harness is executable context-construction code that selects and formats trajectory history before each planner call.
A.1 VLABench Skills and Harnesses
VLABench skill evolution specializes generic manipulation guidance for a frozen VLA actor, while harness evolution reorganizes history and observations around current execution evidence. The evolved artifacts emphasize canonical commands, progress monitoring, compact context, and failure signals.
- Skill evolution: The evolved VLABench skill adapts the planner to a frozen VLA actor by routing task intents to short canonical commands.
- Skill evolution: It preserves exact object and container tokens, emits one primitive per round, checks execution progress, and changes command form after failures.
- Harness evolution: The final harness foregrounds the most recent subtask and execution count while compressing older rounds and treating current images as the source of truth.
- Harness evolution: It exposes repeated commands or long executions without progress and adds a compact scene-inspection procedure.
A.2 ESI-Bench Skills and Harnesses
ESI-Bench skill evolution retains the action-space contract while adding explicit evidence-use policies. Harness evolution replaces full-history recency with sparse visual memory, selected evidence, and deterministic visual aids from official RGB observations.
- Skill evolution: The evolved ESI-Bench skill retains the interaction contract while adding explicit policies for using visual evidence.
- Skill evolution: It prioritizes official reference images, distinguishes counting from mapping evidence, and requests suitable viewpoints for geometric judgments.
- Harness evolution: The final harness builds sparse visual memory and retains selected historical evidence instead of concatenating the full textual history.
- Harness evolution: It adds deterministic crops and pixel-space aids derived only from official RGB observations.
A.3 Prompts Used by the Evolution Pipeline
The evolution pipeline uses staged judging, summarization, skill optimization, and harness optimization, with benchmark-specific diagnostics and fixed interface contracts. VLABench emphasizes manipulation progress and executor compatibility, while ESI-Bench emphasizes evidence acquisition, exploration, answer support, and environment validity.
- Shared pipeline: Both benchmarks use judging, summarization, skill optimization, and harness optimization, while adapting inputs and diagnostic criteria to their interaction interfaces.The shared roles implement the textual-feedback loop described in the method.
- VLABench Prompt Instantiation: VLABench exposes progress through before/after camera observations, generated subtasks, and VLA execution statistics.These inputs support round-level diagnosis during planner execution.
- VLABench Prompt Instantiation: Round-level VLABench judgments assess observable progress, execution and reasoning quality, failure causes, and context effectiveness, while episode summaries preserve actionable cross-round patterns.The summarizer also covers instruction fidelity, repetition, decomposition quality, likely root cause, and context effectiveness.
- VLABench Prompt Instantiation: The VLABench skill optimizer replaces the planner skill while freezing the context code and executor, using aggregate feedback and the VLA actor’s command distribution.The harness optimizer separately replaces build_context under documented multimodal, helper, sandbox, and runtime-contract constraints.
- ESI-Bench Prompt Instantiation: ESI-Bench replaces manipulation progress with active-perception evidence signals, uses exact-match outcomes for scoring, and separates invalid environments from agent failures.This prevents renderer or runner faults from becoming optimization evidence.
- ESI-Bench Prompt Instantiation: Its diagnostic judger audits evidence acquisition, retention, use, exploration cycles, answer support, and confidence, routing dominant failures to the skill, harness, environment, or neither.The batch summarizer aggregates accuracy, category results, validity counts, execution statistics, and repeated evidence-supported behaviors without prescribing edits.
- ESI-Bench Prompt Instantiation: The ESI-Bench skill optimizer maps systematic feedback to evidence-acquisition and decision policies while preserving the official interface and fixing the context harness.Policies include reference use, falsifying viewpoints, recovery, legal-action discipline, and confidence-aware stopping.
- ESI-Bench Prompt Instantiation: The ESI-Bench harness optimizer modifies only context construction, enforcing bounded observable evidence, sandboxed code, required interfaces, artifact validation, and payload limits.Its design space includes visual memory, deterministic RGB-derived crops, trajectory state, evidence ordering, loop recovery, and geometric aids.
A.4 Optimization Cost Accounting
The appendix accounts for API-equivalent evolution costs from logged token usage and documents evolved context harnesses and prompt templates for both benchmarks. The reported complete-run costs are approximately USD 2.25 on VLABench and USD 2.83 on ESI-Bench.
- Cost accounting: Complete evolution runs cost approximately USD 2.25 on VLABench and USD 2.83 on ESI-Bench.These API-equivalent costs are computed from logged input and output token usage.
- Cost accounting: The accounting uses Qwen3.6-27B list prices of CNY 3 per million input tokens and CNY 18 per million output tokens, converted at CNY 7.2 per USD.The resulting approximate rates are USD 0.42 per million input tokens and USD 2.50 per million output tokens.
- Appendix artifacts: The documented appendix artifacts include benchmark-specific evolved skills, context-code harnesses, and judging, summarization, and optimization prompt templates.VLABench artifacts emphasize camera evidence and execution statistics, while ESI-Bench artifacts emphasize evidence, exploration, answer, context, and environment diagnostics.
- Appendix artifacts: The prompt templates expose round-level judgments and terminal outcomes for episode summarization before optimization.This staged representation supports textual feedback rather than direct processing of every raw trajectory by the optimizers.
- Appendix artifacts: The VLABench templates separately optimize a complete replacement planner skill and a complete context-construction implementation under their respective fixed-component contracts.The skill optimizer treats the context builder as read-only, and the harness optimizer treats the selected planner skill as read-only.
- Appendix artifacts: The ESI-Bench optimization templates separately evolve an evidence-acquisition skill and a bounded harness for visual memory, RGB-derived views, evidence routing, and loop recovery.The skill optimizer fixes the context harness, while the harness optimizer fixes the evolved skill.