Efficient Skill Grounding via Code Refactoring with Small Language Models

2026-06-06

AI Summary

Existing reusable skills for embodied agents often fail when faced with changes in body morphology or environment, and struggle to achieve efficient skill grounding without the support of large language models. To address this challenge, this work proposes the RECENT framework, which innovatively decouples skill semantics from execution bindings by representing skills as code. Instead of generating new skills from scratch, RECENT leverages small language models to locally reconstruct execution bindings while preserving the original skill semantics. This approach enables efficient and robust long-horizon skill grounding. Experimental results demonstrate that RECENT significantly outperforms existing small-language-model-based methods across diverse robot morphologies and dynamic environments, achieving performance on par with large language model baselines.

Abstract

Effective skill grounding is essential for deploying reusable skills in embodied agents, as even minor embodiment or environmental differences can render an entire skill incompatible. This challenge is particularly pronounced in embodied settings, where agents must operate in dynamic, partially observable environments without access to large language models (LLMs). In this setting, reliance on LLMs is impractical, while small language models (sLMs) remain insufficient for the effective skill grounding required for reliable long-horizon control. We present RECENT, a refactoring-centric agent framework that enables efficient skill grounding with sLMs by decoupling skill semantics from embodiment- and environment-specific execution binding. By representing skills as executable code, RECENT preserves the semantic intent encoded in a skill's control structure while grounding it by modifying only execution bindings through localized refactoring, rather than regenerating code from scratch. We evaluate RECENT across diverse skill grounding scenarios spanning multiple robot embodiments in dynamic environments, demonstrating robust long-horizon performance when deployed with an sLM. Across all scenarios, RECENT achieves the best performance among sLM-based Code-as-Policies (CaP) methods and matches the task performance of LLM-based CaP.

Problem

• skill grounding
• embodied agents
• small language models
• code refactoring
• long-horizon control

Innovation

• skill grounding
• code refactoring
• small language models
• embodied agents
• Code-as-Policies