Researchers trained a robot's task policy - the model that directs its actions while it completes a task - on short skills that repaired failed runs. In a drawer task, the policy's unaided success rose from 0 of 20 attempts to 7 of 20 after one update.
The method, called skill-space shooting, starts with a task policy and a library of reusable robot skills such as grasping, placing or sweeping. A learned value model watches the scene and the policy's next action chunk for signs of failure. When it flags a risky move, Gemini 2.5 Pro selects a relevant skill from the library and checks whether the physical repair worked.
After an accepted repair, control returns to the task policy. The researchers keep only the successful skill-executed repair segments from completed runs and use them to retrain the policy. At evaluation, they remove the value model, the Gemini judge and the skill library. The test therefore asks whether the task policy learned the correction, rather than whether an assisted system can finish the task.
The team tested four physical tasks, including stacking cubes, putting a tape roll in a drawer, preparing coffee and sweeping blocks. The drawer policy initially failed all 20 evaluation attempts. After one update it succeeded in 7 of 20, while the sparse-reward baseline remained at zero after its first update and reached at most 2 of 20. Each task had one training run; evaluation used 30 attempts for cube stacking and 20 for each other task.
The result is a small robotics experiment, not evidence of unattended self-improvement in the field. People supplied the initial task and skill demonstrations and reset the scenes between rollouts, while the failure detector was trained separately for each task. The study shows how a policy can learn from targeted repairs in these setups; it does not establish that the approach will transfer to new tasks without preparation or run continuously without human resets.