Intel / article

Self-Evolving Embodied Agents via Skill-Harness Evolution

Source: arxiv.org
Author: Peidong Wang, Zhiming Ma, Ying Chang, Xufang Luo, Xiaocui Yang, Shi Feng, Yuqing Yang, Dongsheng Li
Date: Published Aug 13, 2026

Context

SHAPER explores how an embodied can adapt to an environment while leaving its underlying AI model unchanged. It revises reusable skills and the surrounding code that manages and execution, using experience from attempts in the target environment. The same frozen model can both plan actions and propose improvements to this surrounding system. The researchers evaluate that approach on VLABench and ESI-Bench.

The proposed adaptation happens in artifacts the agent uses, rather than in its learned model parameters. That offers a route to improving an agent when retraining is impractical, although it still requires running the agent and evaluating changes in its environment.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.

Key quotes

“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.”
Peidong Wang, Zhiming Ma, Ying Chang, Xufang Luo, Xiaocui Yang, Shi Feng, Yuqing Yang, Dongsheng Li

“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.”
Peidong Wang, Zhiming Ma, Ying Chang, Xufang Luo, Xiaocui Yang, Shi Feng, Yuqing Yang, Dongsheng Li

“In SHAPER, the same frozen model can serve as both planner and optimizer, refining its external skills and context-code harness without parameter updates.”
Peidong Wang, Zhiming Ma, Ying Chang, Xufang Luo, Xiaocui Yang, Shi Feng, Yuqing Yang, Dongsheng Li

“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.”
Peidong Wang, Zhiming Ma, Ying Chang, Xufang Luo, Xiaocui Yang, Shi Feng, Yuqing Yang, Dongsheng Li

Read the source arxiv.org