How are humanoid robots being trained?
This is something I find super interesting and I don't see it discussed much
So from what I understand, there are basically a few approaches companies are using:
Teleoperation + Learning: Someone operates the robot remotely, and the robot learns from watching those movements. Figure has shown this - they have people controlling the robot and it picks up on patterns.
Simulation: Train the robot in a virtual environment first, then transfer to the real world. This is faster and safer but sometimes things don't transfer perfectly.
Learning from videos: This is newer and honestly kind of wild. Tesla mentioned training Optimus partly by having it watch videos of humans doing tasks. The idea is it can learn movements without needing someone to physically demonstrate.
Id be curious to know how long does it take to teach a robot a new task? Like if you wanted it to learn to make coffee, is that a day? A week? Months?
Also, I saw the Sunday announcement where they've been using a glove for training. It seems so obvious and I was suprised to see that this was a very novel approach.
2 Replies
The teleoperation approach is really promising. Figure calls it "learning from demonstration" and it dramatically cuts down training time compared to pure RL.
For your coffee question - simple manipulation tasks can be learned in hours with enough demo data. Complex multi-step tasks with variations probably take weeks to get reliable.
Hours to days is faster than I expected honestly. Makes sense that demo data speeds things up vs learning from scratch.