SHAPER Evolves Skills and a Context-Code Harness for Embodied Agents With Frozen Weights — Aimed at Robots Without Programmable APIs

arXiv (via HuggingFace Daily Papers)·low signalXBlueskyLinkedInCopy link

Peidong Wang, Xufang Luo, Yuqing Yang and Dongsheng Li (arXiv 2608.11350, August 11) introduce SHAPER, which keeps model parameters frozen and instead evolves reusable skills plus a context-code harness through rollouts in the target environment. It specifically targets fixed-interface settings where existing train-free code-centric approaches break because no programmable robot API is available, with results on VLABench and ESI-Bench. A single frozen foundation model acts as both planner and optimizer — the same 'improve the harness, not the weights' thesis as AI4AI at Test-Time, arriving independently in embodied robotics.