OmniXtreme unifies extreme humanoid motion control

OmniXtreme takes direct aim at the "gala problem" in humanoid robotics — the embarrassing gap between impressive product-launch stunts and actual generalizable capability.

• –Two-stage training combines DAgger-based Flow Matching (to merge knowledge from multiple specialist controllers) with actuation-aware fine-tuning that models real motor torque-speed limits and battery undervoltage failure modes
• –157 real-world trials across flips (96%), martial arts (93%), handsprings (88%), and breakdancing (86%) — success rates this high on high-impact dynamic motions are a genuine benchmark advance
• –The power-safety regularization term is the unglamorous engineering win here: preventing motor overcurrent on landing is what actually makes repeated real-world trials possible
• –Enters a crowded generalist-control field alongside NVIDIA SONIC and Amazon's OmniRetarget, but with more extreme motion coverage than either published demo
• –BIGAI is a Beijing government-affiliated institute — this is also a geopolitical signal about China's robotics research ambitions