Hugging Face researchers have unveiled SHAPER, a novel framework that enables embodied agents to self-evolve without any training. The approach keeps the foundation model's parameters frozen and instead optimizes the surrounding system—reusable skills and a context-code harness—through environment rollouts.
Embodied agents are increasingly built as systems around foundation models, where performance depends not only on model weights but also on the skills, context, action interfaces, and execution harness. Traditional adaptation methods like supervised fine-tuning and reinforcement learning require additional data, rewards, and training runs, while many train-free code-centric approaches rely on programmable robot APIs that may be unavailable in fixed-interface settings.
SHAPER addresses these limitations by using the same frozen model as both planner and optimizer. It refines its external skills and context-code harness without any parameter updates, making it a practical alternative when model training is expensive, unavailable, or undesirable.
The framework was evaluated on VLABench and ESI-Bench, covering embodied agents with different low-level action interfaces. Comparisons were made against pure execution, supervised fine-tuning, and test-time-scaling baselines such as verifier-free selection and voting. The results suggest that skill-and-harness optimization is a viable route to self-evolving embodied agents.