Ep#76: OmniXtreme: Breaking the Generality Barrier in High-Dynamic Humanoid Control

In this talk

1. 1:36

   The generality barrier

   Wang describes the gap between varied, difficult motions in character animation and the narrower skills demonstrated by real robots. He calls the loss of tracking quality as motion diversity and difficulty increase the generality barrier. He attributes it to learning limits and the challenges of transferring policies to hardware.

2. 7:16

   Learning a unified controller

   Wang says multi-motion training can make optimization unstable, with conflicting gradients and weaker performance on difficult motions. OmniXtreme trains expert policies on challenging motions, then uses DAgger to distill them into a flow-matching controller. The team reports that flow policies scale better than MLPs as model size increases, and says its controller has about 50 million parameters.

3. 8:22

   Modeling hardware constraints

   The team identifies actuator behavior as a key sim-to-real factor, including torque limits that decrease as joint velocity rises. Its post-training stage freezes the flow policy and trains a lightweight residual network with more realistic actuator modeling, disturbances, and domain randomization. The team also penalizes excessive negative power to reduce current spikes during high-speed landings.

4. 17:01

   Results and deployment limits

   In simulation, OmniXtreme is evaluated on selected extreme motions, FAN-1 data, and sampled unseen motions from AMS. Real-world demonstrations on the Unitree G1 include Cossack dancing, web flips, breakdancing, and handstand walking; inference takes about 10 milliseconds onboard. The authors say they broke five or six robots during development, and that power safety can still limit motions the G1 cannot perform.

5. 38:06

   Open questions for broader control

   The authors say residual post-training is harder than flow-policy pretraining, reporting about 100,000 post-training steps across 16,000 environments. They hope to study generalization and extend the approach to locomotion and manipulation. The discussion points to future challenges such as graceful failure, perception, and maintaining constraints while performing extreme motions.