Physical AI Lab // Episode 1.6 // Quiz
Imitation learning: Learning from demos
Just copy it? Why it's not that simple
01
A behaviour-cloning policy gets the first few steps right, drifts slightly halfway, and then goes further and further wrong. What is the main reason?
Wrong actions had crept into the demonstrations
Off course, it's where no demo went; errors snowball
The control rate is too high to finish each action
Behaviour cloning picks a random action at every step
02
Instead of one action per step, ACT outputs a short chunk of actions at a time. What is this mainly for?
Less jitter and drift; errors snowball less
So the model sees more images each time
Fewer cameras needed when recording demos
So the robot can learn with no demos at all
03
Going around an obstacle, both left and right are correct. A policy trained the old way drives straight into the obstacle. What is the most likely cause?
The demos only recorded going left, never right
It averaged left and right, ending in the middle
It used a diffusion policy, which is too random
The obstacle was outside the camera's field of view
04
The ACT paper reports 80–90% success, with about ten minutes of demonstrations, on real tasks such as opening a condiment cup and slotting a battery. Which inference holds up best?
Any task reaches 80–90% with ten minutes of demos
It is a simulation result and would be far lower on a real robot
It was measured on the real tasks the paper names
Recording a few more demos would surely raise the success rate
05
You already have enough demonstrations for one scene, and your people can do only one more thing. By the paper's conclusion in the episode, what is the best use of them?
Record twice as many demos in the same scene
Keep only the best few and delete the rest
Train on the recorded demos a few more times
Record in more settings, with more objects
Check