SmolVLA, Flow Matching & SO-101 Deployment. Go from architecture to hardware — deploy a vision-language-action model on a real robot arm, debug the 10 critical bugs, and run bimanual inference.
Open Lecture SlidesFrom flow matching math to real-robot deployment — build, debug, and deploy a VLA end-to-end.
Build flow matching from first principles — optimal transport paths, velocity fields, and the connection to diffusion.
Open in ColabUnderstand cross-attention conditioning and compare VLA architecture choices — from self-attention to cross-attention fusion.
Open in ColabPixelShuffle token compression, layer skipping, and async inference — the tricks that make SmolVLA fast enough for real-time.
Open in ColabSimulate the full deployment pipeline — from recording demonstrations to training, evaluating, and running inference on SO-101.
Open in ColabWalk through the 10 critical bugs we hit deploying SmolVLA on SO-101 — with interactive exercises to find and fix each one.
Open in ColabSeven parts covering the full journey from architecture to deployment — jump to any section in the slides.
Real SO-101 bimanual demos from our lab — SmolVLA controlling two arms for box-pass tasks.
Two SO-101 arms performing a coordinated box-pass task with SmolVLA inference.
Handover variation with different object placement and camera angles.
Extended sequence showing recovery behavior and multi-step manipulation.
Language-conditioned pick-and-place with voice control for object selection.