Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Legged Robots Still Trip Over Stairs
Transcript
- Lucas: So Boston Dynamics released a new Atlas demo in late April — humanoid robot, whole choreographed routine in their lab. And it's beautiful. It does a backflip off a small platform, catches itself, and then walks up a flight of stairs like it's nothing. Luna: And I immediately thought — okay, so stairs are solved now, right? We can move on. Lucas: That's exactly the impression they want you to have. But if you watch the uncut footage from their own blog, there's a telling cut. They only show one stair configuration — straight flight, identical riser height, perfect lighting. No spiral stairs, no uneven treads, no landing turn. Luna: So it's a curated test. Which makes sense for a marketing demo, but it doesn't tell us whether the robot actually understands stairs in general. Lucas: Right. And that's the real challenge — generalisation. A human can walk up any reasonably constructed stairway on the first try. A legged robot, even a very advanced one like Atlas, is still brittle outside the exact conditions it was tuned for. Luna: I remember reading about a test a few years back where a robot that could climb a 20-degree ramp perfectly fell over on an 18-degree ramp because the control parameters were hard-coded for 20. Is that still the state of the art? Lucas: It was for a long time. But there's been a real shift in the last two or three years, driven mostly by simulation-based reinforcement learning. Instead of hand-tuning every parameter, you let the robot learn a policy in simulation across thousands of stair variations — different heights, widths, textures, even dynamic changes like a step that shifts slightly under load. Luna: Who's doing that well right now? Lucas: ETH Zurich's Robotics Systems Lab is probably the leader. Their quadruped ANYmal has a system they call 'blind stair walking' — it uses no vision at all. Just proprioception — joint angles, torques, inertial measurement. It walks up and down stairs by feel, essentially. And in their 2025 paper, they demonstrated it on stairs with varying riser heights between 15 and 22 centimetres, which covers most residential stairs worldwide. Luna: Wait — no cameras? So it's purely tactile? Lucas: Purely. The idea is that vision can be unreliable — glare, shadows, fog, debris. If the robot can do it blind, it's a much more robust foundation. Then you layer vision on top for planning. ANYmal's blind policy was trained in simulation over the equivalent of 10,000 hours of stair climbing, using a technique called domain randomisation — they randomise friction, mass, ground stiffness, even sensor noise, so the policy learns to handle real-world mess. Luna: And it works? They've tested it on actual buildings? Lucas: They did a field test in an ETH campus building — a concrete spiral staircase with a tight radius. The robot slipped on the third step, recovered, and completed the ascent. That's the kind of failure that, five years ago, would have been a hard crash. The fact that it recovered mid-stair is the real progress. Luna: So the industry is moving from 'don't fall' to 'fall gracefully and get up'. That's actually a huge conceptual shift. Lucas: It's the difference between a scripted performance and a robust skill. And that's what you need if you want a robot in someone's home. Because home stairs are not lab stairs. They have carpet that bunches, a child's toy on the third step, a pet running underfoot. The robot has to adapt in real time. Luna: Which brings us to the question of bipedal versus quadruped. Most people assume a humanoid is the natural form for stairs because we use them. But quads seem to be much further along on stair competence right now. Lucas: Quads have a lower centre of gravity and four contact points, so the stability margin is much larger. ANYmal can take a misstep and still have three legs on the ground. A biped has zero margin in single-support phase. If the foot placement is off by two centimetres on a stair edge, the robot is falling. Luna: So the biped problem is fundamentally harder. But the payoff is also bigger — a humanoid fits human environments without modification. Lucas: Exactly. And that's why companies like Boston Dynamics and Tesla and Figure are still pushing bipedal. But I think the honest timeline from researchers I've talked to is that a reliable bipedal stair climber — one you'd trust to carry a laundry basket down a flight — is probably five years away at least. And that's with the current pace of simulation learning. Luna: Five years feels both fast and slow. For the home robotics market, that means the first useful robots in houses are probably not going to be humanoids. They'll be quads or wheeled bases that are designed to avoid stairs entirely. Lucas: That's actually a really practical point. A lot of the home robot startups are focusing on single-floor environments, or they build in a 'call the elevator' capability. Stairs are the obstacle that defines the floor plan. Luna: If today's episode was useful to you and you want to keep it ad-free, buy me a coffee dot com slash fexingo helps. Lucas: Yeah, it's a small ask but it genuinely makes a difference. And we appreciate it. Luna: So, back to the stair problem — I want to talk about a specific failure mode that I think is underappreciated. Lucas: Which one? Luna: The transition from stair walking to level ground. Most stair tests start at the bottom, go up, then stop. But what about walking up a flight, turning around on a landing, and walking down a different flight? The landing is a small flat area with a 90-degree turn. Lucas: That's a classic problem in locomotion — it's called the 'stance foot transition'. On a landing, the robot has to change its gait from stair mode to walking mode and then back, while rotating. In the ETH Zurich paper, that was actually the highest failure point. They had a 12 percent failure rate on landings versus about 4 percent on straight stair sections. Luna: So the narrow landing is harder than the stairs themselves. That makes sense — it's a discontinuity in the terrain model. Lucas: Right. And that's where the simulation approach really shines because you can generate millions of landing configurations — different widths, different turn angles, different surface materials — and train the policy to handle all of them. The old method would require a human engineer to write a separate state machine for landing behaviour. Luna: How fast is the simulation training? Can you train a policy overnight? Lucas: With a decent GPU cluster, yes. The ANYmal stair policy took about 12 hours of simulation time using Nvidia's Isaac Gym simulator. That's 12 hours to generate what amounts to centuries of real-world stair experience. Luna: And then you transfer that policy to the real robot and it just works? Lucas: Not automatically. There's a sim to real gap. The robot's sensors have noise that the simulation didn't capture perfectly. The actuators have latency. The floor might be more slippery. The ETH team used a technique called 'system identification' — they run the robot on a simple test motion, measure the discrepancy, and update the simulation parameters. Usually one or two iterations gets them close enough. Luna: So the hard part isn't the learning anymore. It's the bridging. And that bridging is getting faster every year. Lucas: Much faster. In 2020, a sim to real transfer for a walking policy would take weeks of tuning. Now it's down to a couple of days. The compounding effect is that the research cycle is shorter, so more ideas get tested. Luna: Where do you see the biggest open problem right now, then? If sim to real is improving, simulation is cheap, what's the bottleneck? Lucas: I think it's perception under uncertainty. The robot needs to see the stairs, classify the type, estimate the geometry, and then execute the right policy. But what if the stair is partially occluded by a person walking down? Or what if the lighting is so low that the depth camera fails? The blind walking approach avoids that, but blind walking can't handle obstacles on the stairs — like a box or a pet. So you need both modalities working together. Luna: Multi-modal fusion — vision and touch — is still a research frontier for manipulation. For locomotion, it's even less mature. Lucas: Exactly. And that's where I think the next big breakthrough will come. Not a better actuator or a lighter battery, but an algorithm that can switch between vision-based and touch-based control mid-stride, depending on which signal is more reliable at that moment. Luna: So the robot that can handle your laundry down the stairs will need to see a sock, feel the step shift, and adjust its foot — all within a quarter of a second. Lucas: That's the goal. And we're not there yet. But the trajectory is clear — stair climbing is going from a hard-coded stunt to a learned skill. And once it's a skill, it's just a matter of training hours.