THE MECHANISM
ACT, diffusion, and flow matching
03ACT compresses behavior into chunks
→04Diffusion iteratively denoises
→05Flow matching learns a velocity field
FAILURE → REPAIR
Common trap
“The most expressive head is always best.”
Better model
A head that misses the control deadline is not usable. Quality, latency, chunk horizon, and hardware are a coupled choice.
YOUR NEXT GATE
Action-head bake-off
- 1Verify your installed LeRobot CLI with lerobot-train --help.
- 2Train ACT as the fastest path to a working baseline.
- 3Train diffusion on the same dataset as a quality reference.