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

Code-writing large language models (CodeLLMs) generate executable code policies for embodied agents by translating natural language goals and environmental constraints into structured control programs. However, policy generation in open-domain embodied environments suffers from two fundamental limitations: (i) delayed decoding caused by repetitive prefill computation over long prompts, and (ii) limited robustness due to fully generative decoding, which often produces API mismatches, missing safety guards, and unstable control logic. To address these limitations, we present FCGraft, a Functional Cache Grafting framework. FCGraft maintains a library of function-level validated code skeletons and their associated prompt-level Transformer key–value (KV) caches, and synthesizes new policies by retrieving relevant functions and grafting their KV caches when a new task is provided. Given retrieved function caches, FCGraft performs cache grafting via stitching, which composes cached function segments into a composite policy, and patching, which locally adapts only the necessary code regions to satisfy task-specific parameters and constraints with minimal additional decoding. By eliminating redundant prefill computation, this approach reduces generation latency, while reusing validated control structures improves robustness over prompt-level caching methods RAGCache, achieving $18.31\$% higher task success rate and $2.3\times$ faster policy synthesis.

We study how robots can use large language models to write control code more quickly and reliably. Recent robots can translate natural language instructions into executable programs, but they often need to process long prompts again and again for every new task. When the generated code fails, the robot may also need to regenerate the full program, which slows down execution and can make the robot less responsive in dynamic environments. We propose FCGRAFT, a framework that lets the robot reuse previously successful code at the function level. Instead of generating every policy from scratch, FCGRAFT stores validated robot-control functions together with their internal model caches. It then stitches relevant functions together for a new task and patches only the small code region that needs correction when an error occurs. This makes robot policy generation faster while also improving reliability. By reusing validated control structures, FCGRAFT reduces inconsistent code, API mistakes, and missing safety checks, helping embodied agents act more robustly in open-domain and real-world tasks.