Flow Matching for Robot Policies
Integrating flow-matching action heads into the Octo generalist robot policy, generative action modeling for vision-language-action (VLA) systems.
- Python
- PyTorch
- Flow matching
- Robot learning
A research line on how robots should generate actions. Octo is an open generalist robot policy; this work replaces its action head with a flow-matching formulation, a continuous-time generative approach, and studies how that changes what the policy can learn.
Variants
- OctoWithFlowMatching, the core integration of a flow-matching action head into the Octo VLA policy.
- FPO, a flow-policy-optimization variant.
- COT, a conditional optimal transport variant of the flow objective.
The throughline is the same question that runs under most of my work: given a hard objective, what’s the representation that makes it learnable and fast?