NYU's Steertuning adapts robots to new tasks from 5 human demos in under 5 minutes
mangahomanga · x · 2026-11-10
NYU researchers (Lerrel Pinto, Irmak Guzey, et al.) introduced Steertuning, letting robots learn new tasks in new environments from just 5 human demonstrations, with the whole collect-to-deploy process taking under 5 minutes.
- Pretraining: a 400M-parameter diffusion policy is pretrained on 100 hours of densely annotated egocentric human random interactions, making it robust to background changes with spatial and object generalization
- Finetune + steer: given 5 demos of a new task, finetune for only 2 minutes (2000 updates); at deployment, the demo whose observations best match the environment steers the policy's actions during reverse sampling
- With a purpose-built embodiment-transfer data pipeline, it achieves a 71% success rate across 12 tasks in 12 new environments, from labs to kitchens in new homes
Dataset, model weights, and code are open-sourced.
More from Embodied
- TRI's Haruki Nishimura to talk trustworthy deployment of large behavior models — GlenBerseth · 2026-11-10
- First CyberCab rides impress: 'I think it's exceeded Waymo in experience' — carlosdponx · 2026-11-10
- GR00T 2 Turbo runs onboard Jetson Thor, driving humanoid full body in real time — yukez · 2026-11-10
- GR00T 2's X-Adapt: wearable human demos teach robots skills with zero robot practice — yukez · 2026-11-10
- GR00T 2's WAM API lets reasoning agents steer robots by pointing in the camera view — yukez · 2026-11-10
- NVIDIA unveils GR00T 2 world-action model: 40% zero-shot success, human demos transfer to robots — yukez · 2026-11-10