Physical AI Lab // Episode 2.5 // Quiz
Sim2real: domain randomization and system identification
Two ways to patch the seam between simulation and reality
01
With the same push, the simulated cup slides far, but the real cup stops early. What is the most likely cause?
The simulation timestep is too large and diverged
The simulated cup's texture is too plain to be realistic
Friction, mass or delay don't match the real robot
The push is too weak and gets lost in sensor noise
02
In one sentence, which best describes the idea of domain randomization?
Tune the simulation's parameters as close to the real robot as possible
New parameters every episode, so reality is just one more variant
Add a little random noise to every observation
Train on the real robot first, then test in simulation
03
You make the friction range extremely wide, and the policy ends up timid and mediocre everywhere. What does this tell you?
Domain randomization doesn't work; drop it
The range is too wide, so the policy got dull
Too few randomized episodes; run several times more
Friction shouldn't be randomized, only mass
04
During system identification you sweep candidate friction values, but the loss curve is almost flat, with no clear bottom. What is the most likely cause?
The sweep is so dense that it smooths the curve out
A flatter curve means a more accurate identification
You should use the average landing point of many trials
This parameter is hard to identify from this data
05
When the two methods are used together in practice, what is the usual order and division of work?
Identify first to narrow the range, then randomize the rest
Randomize widely first, then identify to fine-tune the policy
Pick one; using both cancels them out
Identify only in simulation, randomize only on the real robot
Check