THE MECHANISM
Vision + language + an action expert
03Reuse semantics, specialize action
→04Action interfaces differ
→05Generalist does not mean embodiment-free
FAILURE → REPAIR
Common trap
“A large VLM plus a linear action layer is automatically a robust VLA.”
Better model
Action representation, temporal chunking, data balance, calibration, and deployment latency are first-class parts of the system.
YOUR NEXT GATE
Dissect SmolVLA before running it
- 1Label the vision, language, state, and action paths in the ANIMA diagram.
- 2Find where action chunk length is configured.
- 3Write down the policy device and expected inference host.