Self-Evolving Embodied Agents via Skill-Harness Evolution

2026-08-11

AI Summary

This work addresses the challenge of enabling embodied agents to efficiently adapt to novel environments when model retraining is prohibitively expensive or infeasible. The authors propose SHAPER, a framework that achieves agent self-evolution without additional data, rewards, or fine-tuning by freezing the parameters of a foundation large model and instead jointly optimizing a non-parametric, reusable skill library and a context-to-code executor. This approach uniquely treats the co-optimization of skills and the executor as an adaptive mechanism, leveraging the frozen large model as a planning and optimization engine while iteratively refining external components through environmental interaction feedback. Experiments demonstrate that SHAPER significantly outperforms baselines—including pure execution, supervised fine-tuning, and test-time scaling—on both VLABench and ESI-Bench, validating its effectiveness and novelty.

Abstract

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.

Problem

• embodied agents
• self-evolving
• train-free adaptation
• fixed-interface settings
• non-parametric system

Innovation

• self-evolving
• embodied agents
• train-free adaptation
• skill-harness evolution
• frozen foundation models