SEAR: Self-Evolving Agentic Robotics
The paper tests whether robot agents are actually retaining useful skills, not just succeeding through more retries.
SEAR evaluates three agent roles: writing control code, acting directly in the loop, and fine-tuning a vision-language-action model. Across seven frontier models, the authors report that longer runs help most on harder tasks, while easier ones can be matched by repeated short attempts. Reusable control code and shared routines transfer better than task-specific fixes, but real-robot tests still show a gap on novel human-designed scenes. Different models improved best through different modes, and a text-only model kept pace with a comparable multimodal one. ArXiv · AI/CL/LG's note
SEAR evaluates three agent roles: writing control code, acting directly in the loop, and fine-tuning a vision-language-action model. Across seven frontier models, the authors report that longer runs help most on harder tasks, while easier ones can be matched by repeated short attempts. Reusable control code and shared routines transfer better than task-specific fixes, but real-robot tests still show a gap on novel human-designed scenes. Different models improved best through different modes, and a text-only model kept pace with a comparable multimodal one. ArXiv · AI/CL/LG's note
score 6