Physical AI Lab // Episode 3.1 // Quiz
Inside VLAs: three models
Vision encoder, language backbone, action head
01
A team wants a dual-arm robot to fold clothes, which needs fast, smooth continuous motion. With OpenVLA the motion lags and comes out jerky. By the video's breakdown, which part should they suspect first?
The action head: token-by-token output struggles with high-frequency, dexterous tasks
The vision encoder: fusing two sets of visual features made it misread the scene
The language backbone: Llama 2 can't understand a complex instruction like folding
The parameter count: 7 billion parameters make inference too slow to keep up
02
Someone says: "π-zero and GR00T N1 both generate continuous actions, so their action heads are exactly the same." By the video, which is more accurate?
Yes, both cut each action dimension into 256 bins and predict them as tokens
Both use flow matching; π-zero adds an expert, GR00T N1 a fast-system DiT
No, π-zero actually uses discrete tokens, and only GR00T N1 generates continuously
Yes, both are fast-and-slow dual systems, just named differently in the papers
03
Why does continuous action generation, such as flow matching, suit fast, fine work like folding clothes better?
Models with these action heads have more parameters, so their motion comes out finer
They don't rely on vision and can move from the language instruction alone
Motion is continuous, so it can output a smooth stretch at once, not token by token
Flow matching produces an action in a single step, with no generation delay at all
04
None of the three models trains its seeing and language parts from scratch; they borrow existing open models. What main reason does the video give?
A model trained from scratch can't connect to an action head, so it can't output actions
Open-source licences require a VLA to be built on an existing large model
Robot demonstration data contains no images, so vision has to be borrowed
Borrowed models already know objects and meaning; from scratch is too slow and costly
05
You test the three models on a new dual-arm robot, and GR00T N1, the one with the fewest parameters, does better than OpenVLA. What is the most reasonable reading?
Fit depends on task and body; on another robot the ranking may change
Smaller models are always stronger, so always pick the smallest
The test probably went wrong, since OpenVLA has more parameters and should win
This proves flow-matching action heads beat token prediction on every task
Check