Teach Robots 1000 Tasks in 24 Hours with Just a Single Demonstration
chris_j_paxton · x · 2026-08-03
A new study published on the cover of Science Robotics introduces MT3 (Multi-Task Trajectory Transfer), an imitation learning method that drastically improves robot learning efficiency.
- Core Mechanism: Decomposes manipulation trajectories into sequential 'alignment' and 'interaction' phases, combined with retrieval-based generalisation.
- Data Efficiency: In the few-demonstrations-per-task regime (<10 demonstrations), decomposition achieves an order of magnitude improvement in data efficiency over single-phase behavioural cloning.
- Strong Generalisation: MT3 learns everyday manipulation tasks from as little as 1 demonstration per task while generalising to previously unseen object instances.
- Scale: This efficiency enabled researchers to teach a robot 1000 distinct tasks in under 24 hours, validated extensively through 5650 real-world rollouts.
More from Embodied
- Robot Weaponization Regulation Lags Behind LLM Safety Standards — alex_verem · 2026-08-03
- Lambda Surgical Raises $6.5M to Build World Models for the Human Body — ditzikow · 2026-08-03
- Robotics expert: Legs will be more robust and safer long-term, but wheels work best now — chris_j_paxton · 2026-08-03
- Masked Visual Actions: Turning Video Models into Robotic World Models — jiqizhixin · 2026-08-03
- ELLIS Institute Launches Physical AI Lab Focused on Robot Learning — orvieto_antonio · 2026-08-03
- Robot muscle material combines damage resistance and self-healing, researcher plans to scale up — MarwaEldiwiny · 2026-08-03