Chapter 16 · AI in science and the physical world
Robots that learn
Robots learn skills by practising in simulation and from human demonstrations, then adapt to the real world.
Chapter 16 · AI in science and the physical world
Robots learn skills by practising in simulation and from human demonstrations, then adapt to the real world.
Try 1: the block slips. Fail.
In a simulator, a robot can try thousands of times for free. Each fall teaches it something. Here a toy arm learns to place blocks gently.
A real arm can try only a few times a minute, it wears out, and a bad try can break things. So most of the practice is done inside a computer simulation, where tries are fast, free and safe.
The real floor is slipperier than the simulator’s usual setting. Can you get a good real-world score?
It is great in simulation but the real floor is different. Add randomness so friction changes during practice. A teaching toy with invented numbers, not a real robot.
Why do robots practise mostly in simulation?
A robot is perfect in simulation but drops the block in real life. What is this called?
How does randomising the simulator help?