Physical AI Lab // Episode 2.6 // Quiz
To the edge: quantization, latency, control rate
A robot can only correct itself as fast as its model runs
01
Reading the image, preprocessing, inference and sending the action take 50 milliseconds in total. Used directly as the controller, at most how many hertz can it run?
50 Hz
200 Hz
1000 Hz
20 Hz
02
The measured average latency is only a few milliseconds, yet the robot occasionally jerks and misses the grasp. Which measurement should you look at?
Measure the average latency a few more times
The tail latency: the single worst case
The model's size after quantization
The model's parameter count
03
After int8 quantization in the notebook, the model is four times smaller but runs slower on this CPU. What is the most reasonable explanation?
Tiny network on a CPU: the overhead outweighs the savings
Quantization failed and the smaller size is an illusion
The official 2–3× speedup figure was simply wrong all along
int8 only speeds things up on a GPU
04
A large model can re-plan only a few times a second (about five in Amazon's measurement), but the low-level loop needs a thousand times a second. What engineering fix does the video give?
A faster chip, so the large model runs every step
Quantize the large model to int8 so it reaches 1 kHz
The large model plans chunks; a fast controller runs them
Slow the low-level loop down to match the large model's rate
05
Why does the video insist on computing on the robot rather than in the cloud?
The cloud has no GPUs for vision models
Round trips add latency, and the network can drop
Actions computed in the cloud come back less precise
The robot's chip is faster than the cloud's
Check