A new robot imitation method learns 1,000 tasks in under 24 hours of demo time

chris_j_paxton · x · 2026-07-30

A Science Robotics paper reports a new imitation-learning method that taught a robot 1,000 distinct everyday manipulation tasks in under 24 hours of human demonstrator time.

The work studies two priors for data efficiency: splitting manipulation into alignment and interaction phases, and using retrieval-based generalization. Across 3,450 real-world rollouts, the authors found that this decomposition gave an order-of-magnitude gain in the few-demonstrations-per-task regime, with retrieval consistently beating monolithic behavioral cloning. The resulting method, MT3, can learn some tasks from a single demo and generalize to unseen object instances.

The paper also notes limits: the gains depend on the task family, and the method was evaluated on real-world rollouts rather than simulated toy settings.

Original post →

More from Embodied

Embodied channel →