RoboParty UFO: Open Unsupervised RL for Humanoid Control
RoboParty Lab open-sourced UFO — short for a Unsupervised reinforcement learning Framework for humanoid cOntrol. The pitch is blunt: tracking-based control still dominates humanoids, unsupervised RL could buy disturbance robustness and skill composition, and the field has been missing shared infrastructure.
What UFO actually ships
From the project page and GitHub:
- Faster Forward-Backward (FB) training: the team reports cutting FB training from multi-day runs down to about 6–8 hours with better performance
- TeCH: a contrastive temporal-distance representation learner so unsupervised RL is not locked to one algorithm family
- Multi-robot path: Unitree G1 is the best-tested route; the gallery also shows Unitree H1, AgiBot X2, and RoboParty’s own RP0/RP1, with per-robot training in roughly 2–3 hours once motion data is adapted
- Real-world hooks: motion tracking, goal reaching, robust teleoperation under external disturbance, and a separate
deploybranch aimed at G1 hardware
The framework expects RobotState motion data already retargeted to the target body. It does not auto-retarget human mocap across morphologies, and checkpoints are not shared across robots with different action or observation dimensions.
Why the kick videos matter (and what they don’t prove)
Viral clips show engineers shoving and kicking a Unitree G1 while the policy recovers — including mid-teleop. That is a vivid stress test of a broad, connected behavior space learned by unsupervised exploration, not proof of production reliability. The open stack is licensed CC BY-NC 4.0 for non-commercial research use, so “open” here means researchers can fork it, not that factories can ship it without a commercial path.
Skill injection is a quieter claim that may age better: UFO’s multi-source data manifests can mix rare motions (the site shows a cartwheel case) without collapsing the base walking distribution — a common failure mode when teams fine-tune agile tricks.
A Human’s Take
I’m excited about the infrastructure angle more than the hazing reels. If unsupervised humanoid RL stays locked inside five labs with private stacks, the field stalls. UFO is messy in the right places — experimental robot bring-up, G1-first deploy, honest “not supported” rows in the docs. That is how smaller teams get a seat at the table.