Physical AI Lab // Episode 1.5 // Quiz
Where does data come from?
Teleoperation, human video, synthesis: three sources
01
A team wants a set of pick-and-place demonstrations for its own arm, with clean motions and exact joint readings. Of the three routes, which fits best?
Download first-person human videos from the web
Generate motions completely at random in simulation
Use a bigger model to make up for the data
Leader–follower teleoperation, recorded by a person
02
Stepping frame by frame through one demonstration in the SO-101 pick-and-place dataset, you want to know what action the robot took in a given frame. What should you look at?
The image from the downward-facing camera
The six joint readings for that frame
The image from the side-view camera
How many frames the whole demonstration has
03
Someone wants to feed cooking videos from a head-mounted camera straight to a robot arm so it learns the motions. Following the episode, what must be solved first?
The videos' resolution is too low and must go up
There are far too few such videos to learn from
Label actions, then map human hands to robot hands
People in the videos move too slowly; speed them up
04
With a method like MimicGen, about 200 human demonstrations were turned into more than 50 thousand. What should you worry about most next?
There are still too few demos to cover the tasks
However many, the synthetic ones may not look real
People still have to re-record each one during generation
Synthetic data only works for teleoperated robots
05
Two teams record the same number of demonstrations. Team one uses the same table and the same cup every time; team two changes tables and objects a lot. Which team would the episode bet on?
Team two: it covers more cases, more varied data
Team one: uniform conditions make cleaner data
Neither: same number, same result
Team one: more takes in one scene, more precise learning
Check