Physical AI Lab // Episode 1.4 // Quiz
A map of simulators
MuJoCo, Isaac, Newton, Genesis: what each does best
01
The same cup-grasp motion is replayed in two engines. The first half almost overlaps, but as soon as the fingers touch the cup the two trajectories drift further apart. What is the most likely cause?
The two engines render images differently
They approximate contact differently
One engine has its gravity parameter set wrong
The two engines loaded different robot models
02
You make the simulation's time step smaller. What is most likely to happen?
It runs faster and is more accurate
Faster, but contacts pass through objects
More accurate, but slower to compute
Same result, just smoother curves
03
You want to study fine contact between fingers and small parts, and compare results with many papers. From the episode's overview, which engine fits best?
MuJoCo: accurate contact, often used as a baseline
Genesis: very new, aiming to unify many kinds of physics
Isaac: realistic images, good for making image data
Newton: built to handle cloth and soft objects too
04
You want to cite “MuJoCo-Warp makes humanoid simulation over 70x faster” in an article. Which wording is most appropriate?
Independent tests confirm every task is 70x faster
A paper measured 70x faster than real-robot training
DeepMind officially promises at least a 70x speed-up
NVIDIA says humanoid simulation is over 70x faster
05
A team trained a grasping policy in the fastest engine with thousands of parallel environments, but on the real robot it often fails. Which closing reminder best explains this?
The engine isn't fast enough; retrain on a faster one
Too many parallel environments confused the policy
Fast isn't accurate; simulation isn't the real world
Simulation is useless; train only on real robots
Check