PARTS concentrates real-world reinforcement learning on the bottleneck subtasks of long-horizon manipulation. Lightweight residual policies adapt a frozen base policy while reusing its reliable behaviors, improving complete-task performance without human action corrections during rollouts.
