WHRG'26 kicks off Saturday — humanoid field finally gets its
What makes this different from the DARPA Robotics Challenge eight years ago? The hardware has caught up to the ambition. Unitree's G1, Fourier's GR-1, Kepler's Forerunner, and at least three Chinese platforms I haven't seen publicly yet — all confirmed for the mobility track. The manipulation track adds a wrinkle: bimanual assembly under time pressure with parts that aren't pre-staged. That's the real test. Anyone can script a pick-and-place when the fixture is bolted down. Doing it when the workpiece shifts 2mm because the robot's own vibration moved the table? That's where the gap between lab and deployment lives.
The organizers published the task board last month. Five mobility events: uneven terrain traverse, 15-degree incline ascent, dynamic obstacle avoidance, narrow passage negotiation, and a "recovery from fall" test that's mandatory — no opt-outs. Manipulation side: gear assembly, cable routing, electrical connector mating, and a surprise task revealed day-of. Scoring weights completion time 40%, success rate 40%, and energy consumption 20%. That last metric is clever. It forces teams to optimize gait efficiency, not just throw compute at the problem.
I've been digging through the technical packets teams submitted. Fourier is running a whole-body MPC controller updated at 1kHz on an onboard Jetson Orin. Unitree went with a learned policy distilled from sim — they're claiming 40% lower compute draw but the sim-to-real gap on loose gravel is still an open question. Kepler's approach is hybrid: classical ZMP walking for flat sections, switching to a learned residual policy for the uneven terrain. Smart architecture, but the switching logic is where it'll fail if they haven't stress-tested the transitions.
The energy metric is going to separate the demos from the products. Most of these platforms run 2-3 hours on a charge doing lab tasks. The WHRG course demands 45 minutes of continuous operation across all five mobility events plus manipulation. Battery management becomes a strategy problem. Do you run the perception stack at 30fps or drop to 10fps on the flat sections to save 15W? That's the kind of systems thinking that doesn't show up in papers.
Streaming starts 9 AM Beijing time Saturday. The organizer's site has a multi-camera feed with telemetry overlay — joint torques, CoM projection, battery draw, all real-time. If you're building legged systems, the telemetry alone is worth watching. You'll see exactly where each controller bleeds energy and where the perception pipeline chokes.
No idea who takes the podium. But this is the first apples-to-apples comparison the field has had since DARPA, and the gap between "walks in a straight line" and "does useful work in messy conditions" is about to get measured in public.