Physical AI Lab // Episode 2.3 // Quiz
Make it understand a sentence
Fine-tuning SmolVLA: from language instructions to actions
01
With the same SO-101 data, why can't ACT follow "put it on the left" while SmolVLA can try?
ACT has too few parameters to hold language
SmolVLA trains for more steps and has seen every phrasing
ACT has no vision network, so it can't see the target
SmolVLA also takes the instruction as input
02
You start fine-tuning and it immediately runs out of GPU memory. What does the video suggest doing first?
Train from scratch instead, to save the base model's memory
Lower the batch size: slower, but it fits on a small GPU
Cut the training steps from 20,000 to 2,000
Turn off the camera inputs and train on joint states only
03
After fine-tuning, "put the cup in the box" works well, but an odd phrasing it has never heard makes it move erratically. What is the most reasonable explanation?
Asynchronous inference is off, so the model can't keep up
Fine-tuning broke the base model, so retrain from scratch
The phrasing is far from what it was trained on
The batch was too small, so it never learned language
04
Which statement about SmolVLA's asynchronous inference matches the video?
It plans the next chunk while acting, at the cost of a server
It shrinks the model, so inference is faster
With it, fine-tuning no longer needs demonstrations
It only helps during fine-tuning and is switched off for deployment
05
In the team's experiment, the SO-100 success rate reached 78.3%. What does this result mainly show?
Pretraining on community data first raises success a lot
SmolVLA succeeds about 80% of the time on any robot
After 20,000 fine-tuning steps, success settles at 78%
Asynchronous inference raised the success rate by about 30%
Check