The Real Story on Humanoid Locomotion Research That the Headlines Missed

The claim that this gaze behaviour "naturally emerges" through reinforcement learning without explicit supervision is, well, it's a strong claim. I called my old colleague at Siemens who works on similar problems and he was skeptical too. But the results are hard to argue with. They're reporting a 1.2 metre gap traversal, which they say is the largest for perceptive humanoid systems. That's a proper jump. Whether it's repeatable in varied conditions remains unclear.

The third paper is the one that caught my attention most. LadderMan is tackling ladder climbing, which is genuinely one of the hardest locomotion problems out there. Sparse footholds, sparse handholds, whole-body coordination where a single mistake means a fall. I've seen industrial climbing robots before (we looked at inspection applications back in 2015 or so) and they were all either tethered for safety or moved so slowly they were basically useless.

What's interesting here is the two-stage pipeline. They use hybrid motion tracking to learn from a single reference motion, basically one example of a human climbing a ladder, then distill multiple "experts" from that into a unified policy. The manipulation component is clever too. They train a separate policy for on-ladder tasks using what they call a dual-agent formulation, which lets the robot do things like grab objects while maintaining its grip on the ladder.

Now here's the thing. All three of these papers are using vision foundation models to handle the sim-to-real gap in perception. This is becoming standard practice, and I think it's actually the most important trend buried in this research. The locomotion algorithms are impressive, sure. But the real breakthrough is that we're finally solving the problem of training in simulation and deploying on real hardware without everything falling apart the moment lighting conditions change.

I don't want to oversell this. These are research papers, not products. The Unitree G1 costs around $16,000 last I checked, which puts it firmly in the "research platform" category rather than anything practical. And none of these papers adequately address what happens when things go wrong. What's the failure mode when the depth perception glitches? When the gait transition happens at the wrong moment?

But I've been watching humanoid research for a long time now, and the pace of progress in the last 18 months is something I haven't seen before. These aren't incremental improvements. The gap between what Boston Dynamics was showing five years ago and what academic labs are publishing now has basically closed. That's worth paying attention to, even if the press releases are still overhyped.

Sarah Williams · 18 hours ago · 5 min