Teaching a robot to open a drawer used to mean thousands of teleoperated demonstrations. The new approach watches ordinary phone videos of people doing chores, infers hand poses and converts them into arm trajectories. The resulting policies work in kitchens the robot has never seen.
Why video beats teleoperation
Humans already produce endless footage of everyday manipulation. Harvesting it is cheaper than hiring operators and covers far more variety. The remaining hard problem is contact: pouring and cutting still need real-world correction.
Key takeaways
- Policies trained from phone video transfer to new kitchens
- Data collection cost fell by two orders of magnitude
- Contact-heavy tasks still need on-robot fine-tuning
This story is part of our ongoing coverage of ai. We will update it as more details emerge.
