Self-Evolving Embodied Agents via Skill-Harness Evolution
SHAPER adapts embodied agents by evolving their skills and execution harness while leaving the model frozen.
The paper frames agent performance as more than model weights: skills, context, action interfaces, and the surrounding harness matter too. SHAPER uses target-environment rollouts to refine reusable skills and context-code without supervised fine-tuning, reinforcement learning, or parameter updates. The same frozen model acts as planner and optimizer. The authors evaluate it on VLABench and ESI-Bench against execution-only, fine-tuning, and test-time-scaling baselines. HF Daily Papers' note
The paper frames agent performance as more than model weights: skills, context, action interfaces, and the surrounding harness matter too. SHAPER uses target-environment rollouts to refine reusable skills and context-code without supervised fine-tuning, reinforcement learning, or parameter updates. The same frozen model acts as planner and optimizer. The authors evaluate it on VLABench and ESI-Bench against execution-only, fine-tuning, and test-time-scaling baselines. HF Daily Papers' note
score 5