Physical AI Lab // Episode 2.2 // Quiz
Training an arm with data
LeRobot + real SO-series arm data: turning demonstrations into an ACT policy that grasps
01
Why does this episode teach the robot to grasp from demonstrations rather than by writing rules?
Rules could be written, but they would run too slowly
Situations vary too much to write every rule
Rules only work in simulation, not on a real robot
Writing rules needs a trained model first
02
While recording demonstrations, the camera frames were logged half a beat after the joint actions. What is most likely if you train on this data?
Training stops with a format error
Training is a bit slower, but the result is unaffected
Actions get paired with the wrong frames
Little effect, because action chunking fills in the half beat
03
If you make the short stretch of actions that ACT outputs at once very long, what problem is most likely?
It falls back to step-by-step prediction and small errors snowball
The parameter count multiplies and no longer fits in memory
The demonstrations are no longer enough; you must record several times more
It can't keep up with changes and reacts slowly to new situations
04
On the real robot you tried once and the policy picked up the cup cleanly. Why can't you report a 100% success rate yet?
One success may be luck; run dozens of trials and average
Only success rates measured in simulation count
Training didn't reach 100k steps, so the result doesn't count
Success rates can only be computed on a standard benchmark
05
You have no real robot but want to score this ACT training pipeline. What does the video suggest?
Use a task with a simulator and a standard benchmark
Test this policy in last episode's MuJoCo scene
Use the final training loss as the success rate
Replay it on the training demos and check the actions match
Check