Skill-Space Shooting for Autonomous Robot Policy Improvement
The paper proposes using reusable short robot skills as corrective supervision when autonomous policies fail.
Foundation models guide the robot through “skill-space shooting,” exploring possible corrections made from learned behaviors. Successful trials are then folded back into the policy so it can improve without a human demonstration for each fix. The authors report real-world experiments showing repeated autonomous policy improvement, with shared skills reducing teaching needed on new tasks. Source: ArXiv · AI/CL/LG's note.
Foundation models guide the robot through “skill-space shooting,” exploring possible corrections made from learned behaviors. Successful trials are then folded back into the policy so it can improve without a human demonstration for each fix. The authors report real-world experiments showing repeated autonomous policy improvement, with shared skills reducing teaching needed on new tasks. Source: ArXiv · AI/CL/LG's note.
score 4