Humanoid Ice Skating

Can a humanoid learn to ice-skate without ever being shown how? This project trains a Unitree G1 on passive 3 mm knife-edge blades with PPO in MuJoCo, using no motion capture and no imitation data. There is no reward term that says “look like a skater” — the skating gait emerges from the interaction between the contact physics and a task reward.

The key idea: let the physics do the talking

Real ice skating exists because a blade is directionally selective: nearly frictionless along its length, and strongly resistant sideways. The simulation captures exactly this with anisotropic blade–ice contact — elliptic friction cones aligned with each blade — plus a reward that separates the two directions:

• Forward glide along the blade is rewarded.
• Lateral slip across the blade is penalized — and the penalty is ungated (always active), which closed a family of reward-hacking exploits where the policy “walked” on the blades instead of gliding.

Given only this, PPO discovers push-off, glide, and weight transfer on its own.

Engineered for hardware transfer

The training setup is built so the policy has a path to the real robot:

• Asymmetric actor–critic — the critic sees privileged blade-velocity information; the actor only gets observations available onboard.
• Staged speed curriculum from 0.15 m/s up to 1.2 m/s commanded speed.
• Domain randomization over PD gains, link inertia, payload, actuator delay, and blade mounting geometry.

Current results (simulation)

Active project (started July 2026) — hardware transfer is the next milestone.