Functional Cache Grafting: Robust and Rapid Code-Policy Synthesis for Embodied Agents

The article introduces FCGraft, a Functional Cache Grafting framework designed to enhance code generation for embodied agents by addressing delays and robustness issues in policy synthesis. FCGraft utilizes a library of validated code skeletons and their Transformer key-value caches to efficiently compose new policies through a process of cache grafting, which includes both stitching and patching techniques. This method significantly reduces generation latency and improves task success rates by 18.31% compared to existing methods, making it a valuable tool for practitioners developing with LLMs in open-domain environments.