One successful demo becomes an editable skeleton: 92.7% average success across 15 Meta-World tasks

Robot policy learning no longer has to start from large teleoperated datasets: one successful trajectory can be rewritten for new object placements and serve both as a task policy and as a demonstration generator. Earlier options split two ways: simple perturbation-based augmentation is easy to apply but can produce invalid trajectories when the target configuration deviates far from the source, while structured methods such as MimicGen rely on manually specified task structure and trajectory-generation procedures. On 15 Meta-World tasks the average best success rate is 92.7%, with the best checkpoint reached after 53.7k environment interactions on average; on 9 DexMimicGen tasks the trajectory editor succeeds on 98.4% of edits on average.

The edits act on the existing successful motion rather than on low-level robot actions, and they come down to two structured decisions: when a segment ends and how that segment is rigidly corrected relative to its object. The source demonstration is represented as an ordered sequence of object-centric stages, each tied to an anchor object or frame; at each stage start the policy predicts a duration and a residual rigid-body transformation, with the anchor sequence providing only a coarse initialization. These decisions are optimized with PPO using only sparse task-success rewards; on DexMimicGen, BC-RNN policies trained on the generated data reach 99.6% average success at matched dataset scale, against 76.3% reported for DexMimicGen data.

The method targets spatial task configurations where object geometry and task requirements stay fixed and only task-relevant object poses vary, covering single-arm, bimanual, and dexterous manipulation, and evaluated on Meta-World, DexMimicGen, and RoboTwin. The authors propose extending it to richer perceptual inputs such as images and point clouds, to within-segment closed-loop correction, and to changes in interaction mode.

The authors note that the method makes only a small number of high-level decisions at segment boundaries, while the action sequence within each segment does not change in response to new states or observations, so gains are more modest on tasks such as Soccer and Stick Pull that benefit from continual state-dependent correction. On RoboTwin, some tasks require multiple behavior modes, and the current approach adapts within a fixed interaction mode, requiring a separate editor per mode.