SOLO Walks a Humanoid 1.5 km on One Chest Camera
A paper posted to arXiv on 27 August 2026 describes SOLO, a perceptive humanoid locomotion stack that the authors deploy zero-shot with a single chest-mounted depth camera and proprioception. No motion capture, no external odometry, no external map. The same policy, running at 50 Hz on 25 controlled joints, finishes a continuous 1.5-kilometer outdoor route and an indoor mixed-terrain course.
The authors’ diagnosis is blunt. Dense terrain reconstructors smear the edges a foot actually needs, and copying a teacher at the current step does not punish actions that blow up later. SOLO splits the fix in two.
Query maps and future blame
Query Reconstructor (QR) asks a Fourier-encoded query for every cell in a 16 × 32 height map and pulls evidence from a shared depth-plus-proprioception token memory. The paper says QR cuts height-map L1 error by 3.3–4.0× versus dense reconstructors.
Trajectory-Aware MSE (TA-MSE) distillation adds next-state teacher–student disagreement to the PPO reward, so Generalized Advantage Estimation can push that penalty back onto earlier actions.
On stress-test terrains, SOLO reports 97.5% mean traversal success and 96% stepping-stone success. Dense-reconstructor variants sit at 75.0–75.6% mean success and 0–3% on stepping stones.
The project page lists the indoor course as stairs, sparse footholds, a gap, a descent, and a movable obstacle in one continuous run, plus a multi-flight fire-stair ascent.
The same stack on Tiangong Omni
The authors say the locomotion technology in this work was also run on Tiangong Omni robots at the 2026 World Humanoid Robot Games in Beijing: gold and bronze in the 400 m obstacle course, silver in the 100 m obstacle course. At WRC 2026, Omni walked stepping stones and stairs in front of a public audience. The still on the project page, credited to People.cn, is that WRC demo, not a lab mocap stage.
That is a competition result on a named platform, not a claim that SOLO itself took the medals. The paper’s own hardware claim is the chest-camera, 1.5 km route.
A Human’s Take
A kilometer and a half with one noisy chest camera is the sentence that matters. Obstacle-course medals are fun; compounding error on mixed ground is the actual job. I want to see the 1.5 km film without cuts, and I want to know what happens when the depth camera hits sun glare or rain. If QR really keeps stair edges that dense maps blur, that is a small geometric trick with a large walking bill.