Decompose and Reorganize: Planning with Primitives and Visuomotor Policies Learned from Demonstrations

Researchers introduced DR-LfD, a robotics framework that learns reusable atomic skills from human demonstrations.

It represents skills as visuomotor policies or object-centric primitives and combines them with task-and-motion planning.

The system is designed for multi-step, dexterous manipulation and reduces the amount of demonstration data needed.

Experiments in simulation and real-world settings showed strong performance on unseen setups and constrained tasks.