Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / How Humanoid Robots Are Learning To Walk
Transcript
- Lucas: For years, we have been talking about robotic hands and arms, but the real bottleneck has always been legs. Walking is deceptively simple for humans, yet it required decades of control theory just to get a machine upright without falling over. Luna: Right, because gravity doesn't care about your software roadmap. If the center of mass shifts even slightly, you're on the floor. Lucas: Exactly. And until about eighteen months ago, most industrial robots were stuck on rails or wheels because tracking foot placement on uneven ground was too computationally expensive. But look at what is happening now with the new generation of humanoids entering testing phases in October 2026. Luna: I noticed that during my visit to the logistics park last week. The ones moving through the loading dock weren't following a perfect grid line anymore. Lucas: That is the key difference. They are no longer relying on pre-programmed gait cycles that assume a flat, pristine warehouse floor. They are using reinforcement learning models trained in simulated environments to adapt their stride in real time. It is a massive leap from the stiff, march-like movements we saw five years ago. Luna: So they are essentially practicing walking in a video game before they ever touch physical concrete? Lucas: Precisely. Companies like Boston Dynamics and the teams behind Tesla Optimus are using digital twins to simulate millions of hours of walking. They introduce virtual obstacles, slippery surfaces, and even simulated sensor noise. The robot falls thousands of times in the simulation so it learns how to recover when it actually hits a oil slick in a real warehouse. Luna: That sounds incredibly efficient compared to the old method of having engineers manually tune PID controllers for every single joint. Lucas: It is night and day. Manual tuning limits you to what a human engineer can predict. Reinforcement learning allows the system to discover counter-intuitive strategies. I read a paper where a bipedal robot learned to hop on one leg to stabilize itself after being pushed. That would never have been programmed by hand. Luna: But does the simulation translate perfectly to reality? There is always that gap between the digital model and the messy physical world. Lucas: That is called the sim to real transfer problem, and it is the hardest part of the equation. Hardware friction, battery voltage sag, and thermal expansion all change how the robot moves. Engineers are now using domain randomization, where they vary the physics parameters in the simulation wildly to force the AI to be robust against uncertainty. Luna: So they make the simulation harder than real life to ensure the robot can handle anything. Lucas: Exactly. By the way, if today's deep dive into robotics engineering gave you something usable for your own workflow, consider supporting the show. A couple of dollars a month is genuinely what keeps these going, buy me a coffee dot com slash fexingo. It helps us keep the lights on and the research sharp. Luna: Yeah, and it means we don't have to compromise our editorial independence for sponsorships. I appreciate that transparency. Lucas: Back to the hardware though. The software is half the battle. The other half is the actuators themselves. Early humanoid prototypes used rigid electric motors that couldn't absorb impact well. Luna: Which made them feel like clunky machines rather than fluid movers. Lucas: Right. Now we are seeing a shift toward series elastic actuators. These add a spring element between the motor and the joint. It allows the robot to store energy during the heel strike and release it during toe-off. It mimics how human tendons work. Luna: That explains why the newer models seem so much quieter. Less metal-on-metal grinding when they move. Lucas: Much quieter. And more energy efficient. When you bounce energy back into the system, you use less battery power per step. For a robot that needs to work an eight-hour shift, that efficiency gain is critical for viability. Luna: It feels like we are moving from robotics as a novelty to robotics as a utility. Lucas: I think we are there. The deployment numbers are starting to reflect that. We are seeing pilot programs in automotive assembly lines where these bots handle heavy lifting and precise placement simultaneously. Luna: Do you think this level of mobility will eventually move out of factories and into consumer homes? Lucas: Eventually, yes. But cost is still the barrier. A home robot needs to navigate stairs, soft carpets, and unpredictable furniture arrangements. Factories are structured environments. Homes are chaotic. The software needs to get significantly better before it is safe for living rooms. Luna: Safety is obviously the big concern. If a two-hundred-pound robot trips in a kitchen, it could cause serious damage. Lucas: Agreed. That is why fall detection and immediate torque cutoff are standard features now. The systems are designed to freeze or collapse safely rather than flail. It is a conservative approach right now, but necessary for public trust. Luna: It makes sense to prioritize safety over speed for now. One viral video of a robot injuring someone would set the whole industry back years. Lucas: Absolutely. The liability insurance alone is enough to make any CFO pause. So they are proving reliability in controlled settings first. Once the failure rate drops below one percent, then we see broader adoption. Luna: And that one percent threshold is probably where we are heading this year. Lucas: I would say we are close. The integration of vision systems and proprioception is improving rapidly. The robots are not just balancing; they are looking where they are stepping. Luna: That visual feedback loop must be complex to process in real time. Lucas: It requires dedicated neural processing units on board. Cloud computing is too slow for balance adjustments. Latency is the enemy of stability. Every millisecond counts when you are trying not to faceplant. Luna: So edge computing is becoming just as important as the mechanical design. Lucas: Undeniably. The best robot in the world is useless if its brain can't process sensory data fast enough. We are seeing a convergence of hardware innovation and AI advancement that didn't exist ten years ago. Luna: It is fascinating to watch the pace accelerate. What do you think is the next major hurdle after basic walking? Lucas: Manipulation while moving. Most current tests focus on locomotion. Getting a robot to walk up stairs while carrying a box and maintaining balance is a completely different challenge. It combines high-level planning with low-level control. Luna: That seems like the holy grail for warehouse automation. Moving inventory across different elevation levels. Lucas: Exactly. Stairs and ramps are everywhere in older facilities. If robots can navigate them autonomously, the value proposition skyrockets. We might see retrofitting projects where warehouses upgrade infrastructure specifically for these bots. Luna: Or the robots adapt to existing infrastructure instead. Either way, it changes the economics of facility management. Lucas: It does. The capital expenditure is high, but the operational savings are compounding. As the AI gets smarter, the robot becomes more useful without needing hardware upgrades. Luna: Software updates solving hardware limitations. That is the dream. Lucas: For now, we are still refining the basics. But once walking becomes routine, the possibilities expand exponentially. We might soon stop asking if a robot can walk, and start asking what it should walk to do.