Humanoids Are Finally Learning to Pick Themselves Up. About Time.

That's clever. I called my old colleague at Siemens about this and he pointed out that terrain classifiers fail all the time in the real world. Dust on a sensor, weird lighting, a surface that's sort of gravel but sort of not. If your recovery depends on correctly identifying the ground type first, you've added a failure mode.

The Second Approach: Model-Assisted RL

The other paper (arXiv) takes a different angle. They're focused on what they call "sparse footholds," basically stepping stones with gaps between them. Think of a robot navigating a construction site with debris, or honestly, just a messy warehouse floor.

Their trick is combining model-based trajectory planning with reinforcement learning. The model gives you a safe reference path, the RL figures out how to actually execute it when things go sideways. They claim this improves sample efficiency, which, look, matters a lot when you're trying to train these systems without burning through millions of simulation hours.

I don't have access to their full results beyond the abstract, so I can't tell you exactly how well it works compared to pure model-free approaches. The claim is "comparable performance" on stepping stones with "smoother locomotion." We'll see. Smoother is good. Comparable is, well, it's a starting point.

What This Actually Means for Deployment

Here's the thing. Both papers validated on the Unitree G1, which is a 29-DoF humanoid. That's a lot of joints to coordinate during a fall recovery. The fact that they're getting stable results on hardware (not just simulation) is genuinely impressive. When I started in this industry, simulation-to-real transfer was basically a prayer. You'd tune everything perfectly in software and then watch the physical robot do something completely different.

But I want to be clear about what we don't know yet. These are research demos. The papers don't tell us about long-term reliability, how many falls the robot can take before something breaks, or what happens when the terrain is wet. (Wet gravel is a nightmare, I've seen mobile platforms that handled dry gravel fine just slide right off a slope when it rained.)

It's also too early to say whether these approaches will scale to heavier humanoids or different form factors. The G1 is relatively light. A larger robot falling on gravel is a different physics problem entirely.

The Bigger Picture

Look, I've been skeptical of humanoids for most of my career. Industrial arms do the job. Mobile platforms with wheels work. Why add the complexity of legs?

But if you want robots in truly unstructured environments, environments built for humans, you eventually need something that can navigate stairs, step over obstacles, and yes, get back up when it falls. These papers suggest we're actually making progress on that last part.

I'm not ready to say humanoids are ready for prime time. But I'm less skeptical than I was last month. And for an old Kuka guy, that's something.

Sarah Williams · 25 Jun · 7 min