NEEDLE is an offline data-augmentation method that composes useful behaviors from imperfect robot demonstrations. It identifies valuable trajectory segments that lack connections and generates short, verified action bridges between them, using only RGB images and proprioception. This bypasses suboptimal detours, combines complementary successes, and recovers useful behavior from failures, all without environment interaction or privileged state.
The system trains a goal-conditioned diffusion inverse dynamics model to propose actions and a verifier to check whether the actions reach the target. When integrated into policy training, accepted bridges add new action sequences alongside original demonstrations, preserving coverage without synthesizing images. On real-robot tasks, NEEDLE improves average success rate by 21 percentage points over the strongest baseline per task.