Physical AI Lab // Episode 3.5 // Quiz
GR00T fine-tuning in practice
Tuning an open humanoid model to your own robot
01
Why is fine-tuning GR00T so much less work than training a model from scratch?
It has only 3 billion parameters, so one GPU can train it from scratch
Pre-training already taught it a lot; you only train a bit more on your data
Fine-tuning needs none of your own data; the repo's sample is enough
Fine-tuning freezes every weight, so you only edit the modality config
02
Training loss falls smoothly with no errors, but in the open-loop evaluation one joint's predicted curve never matches the real one, as if it were wired to a different joint. What is the most likely cause?
The learning rate is too high and training has quietly diverged
The model access the repo requires wasn't granted, so loading went wrong
The modality config is wrong, so actions map to the wrong joints
The sample data is too small for the model to learn that joint yet
03
After fine-tuning, the open-loop evaluation shows the predicted and true lines close together and a low mean squared error. What's the most sensible next step?
Declare real-robot deployment a success and ship it in a product
Run open-loop a few more times; a low enough error means the real robot will succeed
Switch to TensorRT first; once it's faster, closed-loop will be steady by itself
Run closed-loop on a robot: simulation first, real robot once steady
04
You change the data preprocessing. On the same data, the new version scores 3 percentage points higher than the old, with one run each. Following the video, how should you conclude?
Repeat over more episodes and seeds to tell a gain from jitter
3 points is a clear gain; switch to the new version now
Different results from two runs mean a bug; debug the code first
A single run is fairest; more runs only bring in more noise
05
NVIDIA says that going from one thousand to twenty thousand hours of human video more than doubled the average task completion rate. Which reading is the safest?
Going on to forty thousand hours will surely double the completion rate again
Video grew 20 times, so the completion rate grew 20 times as well
NVIDIA's claim for this range: more human video, over twice the completion
The trend is really about having more parameters, not about data volume
Check