Efficient Skill Grounding via Code Refactoring with Small Language Models

Researchers introduce RECENT, a framework that enables small language models to effectively ground robot skills through code refactoring rather than full regeneration. By decoupling skill semantics from embodiment-specific details, the approach matches LLM-based performance while remaining practical for resource-constrained embodied agents.

RECENT addresses a fundamental challenge in robotics: adapting pre-trained skills across different robot embodiments and environments without access to large language models. Traditional approaches either require expensive LLM inference or fail when small language models attempt to regenerate code entirely, losing semantic structure in the process. This research presents a pragmatic alternative by treating skills as modular, executable code where only environment-specific bindings need modification.

The framework's innovation lies in its refactoring-centric approach, which preserves the control logic and semantic intent of original skills while updating only the execution parameters. This contrasts with regenerative methods that risk introducing errors and semantic drift. The decoupling strategy enables deployment on edge devices and embodied systems where computational constraints make LLM reliance impractical, addressing a growing need as robotics applications scale beyond controlled laboratory settings.

For the robotics and embodied AI industry, RECENT demonstrates that practical skill transfer doesn't require frontier-scale models. This has significant implications for cost reduction in robotic system deployment, particularly for organizations operating multiple robot platforms. The ability to match LLM performance while using smaller models suggests a shift toward more efficient AI-robotics architectures.

The research validates its claims across diverse robot embodiments in dynamic environments, establishing credibility for real-world applications. Future developments should focus on scaling RECENT to more complex skill hierarchies and investigating how the framework performs with increasingly heterogeneous robot platforms. The work opens pathways for developing practical, efficient robotics systems that operate independently of cloud infrastructure.

• →RECENT enables skill grounding for embodied agents using small language models through code refactoring rather than full regeneration.
• →The framework decouples skill semantics from embodiment-specific execution bindings, preserving control logic while adapting to new environments.
• →Performance matches LLM-based approaches while maintaining computational efficiency suitable for resource-constrained robotic systems.
• →Refactoring-centric design reduces errors and semantic drift compared to regenerative skill adaptation methods.
• →Research validates effectiveness across multiple robot embodiments in dynamic, partially observable environments.