SEAR study tests whether robot agents learn from practice or just try more often

SEAR’s researchers say they set out to test whether rising success rates reflect learning or repeated attempts, and whether saved skills and memories help on new tasks, in another simulator or on a real robot. They say the agent can write code that becomes a robot’s policy, act live in the control loop or train its own VLA. The study uses a matched control for every comparison and a sealed judge on unseen scenes.