You can efficiently adapt complex robotic systems to new tasks in the real world by learning small corrections on top of existing policies, combined with strategic use of human feedback to guide exploration—no need to retrain from scratch.
This paper presents HiRe-MoMa, a framework for efficiently fine-tuning mobile manipulation robots in the real world using minimal human guidance. The approach uses a hierarchical residual architecture that learns corrections on top of a frozen base policy, while splitting control between the mobile base and arm to reduce interference.