What's Your Problem? Teaching Robots How to Do Everything
Jun 5, 2025
Chelsea Finn, a Stanford professor and co-founder of Physical Intelligence, dives into the fascinating world of teaching robots to perform everyday tasks. She discusses the current limitations of AI in manipulating objects, even for simple chores like laundry folding. Chelsea emphasizes the complexities of training AI models, mentioning a case study on the Pi 0.5 model for household tasks. The conversation also covers the delicate balance between generalist and specialized robots, the ethics of robotics, and the importance of tactile feedback in robot design.
AI Snips
Chapters
Transcript
Episode notes
Teleoperated Data Collection
- Operators use teleoperation to control robots with lightweight arms like puppets.
- Teleoperation captures synchronized video and motor commands to create training data.
Robot Peeling a Hard-Boiled Egg
- Chelsea Finn peeled a hard-boiled egg with a robot gripper, a difficult task.
- Visual feedback replaces tactile sensing to carefully crack and peel the egg shell.
Efficient Robot Learning Strategy
- Robots can learn tasks like putting on a shoe or tearing tape in less than an hour.
- Data efficiency depends on task simplicity and consistent conditions during demonstrations.