Legged Robots: How Walking Went From an Engineering Problem to a Learning Problem

Ten years ago, making a robot walk across a flat floor took a team of engineers and years of tuning. Today, a policy trained in simulation can be dropped onto real hardware and just work. On the RobTalk podcast, Felix Frank, Tech Lead in RobCo's Robot Intelligence team, walked through what actually changed. He spent about six years on robotics research at Volkswagen, then worked on whole-body control at a US humanoid robotics startup, and now focuses on manipulation. This article follows the arc he describes: how legged robotics moved from hand-engineered control stacks to learned behavior, and what that shift means for factories.

Key Takeaways

• Walking used to be a modeling problem: Classical control split locomotion into footstep planning, state estimation, and a large online optimization, each subproblem hand-tuned separately.
• The bottleneck was the unexpected: Anything the robot had to handle needed to be inside the plan beforehand, which made contact with doors, walls, or objects extremely hard to model.
• Reinforcement learning moved the work upstream: Instead of tuning controllers online, engineers now tune simulations and reward functions, then transfer the trained policy to hardware.
• Domain randomization makes the transfer hold: Varying physics, friction, and mass distribution in simulation produces policies robust enough to run on real hardware with no real-world training data.
• The methods carry over: The same simulation-to-reality pipeline behind legged robots is what makes learned manipulation viable on the factory floor.

What Makes Legged Robots So Hard to Control?

A legged robot has to stay upright while continuously breaking and re-establishing contact with the ground. Frank describes walking as a process most people never think about consciously: "you can kind of think of it as controlled falling. You're leaning forward, you're falling forward and then you're catching yourself with your feet."

Do that in the right rhythm and the contact forces at your feet absorb gravity. Get the timing wrong and you fall. A wheeled robot never faces this problem, because its contact with the ground is continuous and predictable. A legged robot renegotiates that contact with every step.

The degrees of freedom involved make it harder still. A humanoid has dozens of joints, all coupled. Move one and the mass distribution of the whole system shifts.

How Did Engineers Make Robots Walk Before Machine Learning?

The classical approach broke locomotion into subproblems, solved each one, and reassembled them into a control stack. Frank lays out the pieces:

• Footstep planning: Given a terrain and a direction, decide where each foot lands.
• Center-of-mass planning: Keep the body's mass centered inside the support area those footsteps create.
• State estimation: Determine where the robot actually is right now, fusing inertial sensors with joint encoder data.
• Online optimization: Solve the whole balance problem continuously, at rates up to one kilohertz.

State estimation alone was a substantial engineering effort. Inertial measurement units, the same sensor class in every smartphone, are noisy. To use them, the system first has to know which foot is currently planted. Only then can it treat that foot as a fixed reference and reconstruct everything else through the kinematic tree of joint angles. That requirement drove hardware cost. As Frank notes, most robots needed expensive force sensors in the feet, just to tell whether they were firmly planted or slipping.

Done well, the reconstruction yields something close to a digital twin of the robot, derived entirely from sensor data.

Where the Classical Approach Hit Its Limits

The problem was not accuracy. It was brittleness. As Frank puts it: "all of this needed to work flawlessly together and any little disturbance that you couldn't plan for would basically break the system."

Worse, anything the robot might do had to be anticipated in the plan. Walking across open floor is a clean case, with two feet and one flat surface as the only contacts. Add a third contact and the math changes. Opening a door, leaning on a wall, or picking up an object introduces external forces that propagate from the hand through the entire body. Each of those cases had to be modeled and compensated for individually.

Humans handle this without thinking. As Frank observes, we spend countless hours as children playing and trying things out. Reproducing that as a set of hand-tuned equations proved cumbersome.

That does not mean the approach failed. Boston Dynamics' Atlas robot ran parkour courses and performed backflips years ago, built on exactly this methodology. It simply took an enormous amount of engineering to get there, and progress eventually plateaued.

What Changed With Reinforcement Learning?

The field moved from solving subproblems by hand to learning entire behaviors from data. Podcast host Clemens draws a parallel to natural language processing. Syntax, semantics, entity recognition, and machine translation were once separate research streams. Large language models absorbed them all.

Frank places the shift roughly: quadrupeds around 2019 to 2020, humanoids later, closer to 2023 or 2024. Two enablers made it possible.

The first was affordable hardware. Twenty years ago, humanoid research required labs to build custom machines that were expensive and fragile. MIT's Cheetah program pushed quadruped design forward, and affordable platforms followed. For humanoids, Unitree's G1 turned research robots into something closer to an off-the-shelf tool.

The second was GPU-parallelized simulation. Reinforcement learning is not data-efficient, so it only becomes practical when a simulator can generate enormous volumes of experience quickly. NVIDIA's Isaac Gym, released in 2021, made this standard. The follow-up work "Learning to Walk in Minutes" trained locomotion policies in minutes rather than days.

Training in Simulation With Domain Randomization

A policy trained in a perfect simulation would fail immediately on real hardware, because no simulation is perfect. Domain randomization solves this by making the simulation deliberately inconsistent.

During training, the system varies mass distribution, friction coefficients, and small deviations in kinematic structure. Frank explains the reasoning plainly: "you're not going to get the same robot twice." Two units off the same production line differ slightly, and so does every floor they walk on.

Randomization also covers active disturbance. The robot "would basically be trained while constantly being pushed around and shoved around and it would have to be robust to all of this." A policy that has absorbed millions of simulated shoves handles a real stumble it has never encountered.

Sim-to-Real Transfer Without Real-World Data

The result is zero-shot transfer: a policy trained entirely in simulation, deployed to hardware without ever having seen real-world data. It works because the underlying physics is Newtonian and well understood, and because randomization forces the policy to cover the range where simulation and reality diverge.

What this does not do is eliminate engineering work. It relocates it. In Frank's words, "the engineering problem kind of went away from we hand tune all these algorithms for the sub problems to we tune the simulation, we tune the reward functions."

The deeper change is architectural. The old stack ran a heavy optimization continuously while the robot moved. The new one does that work in advance: "this complicated optimization problem gets amortized into a function approximator." The neural network outputs a solution without running an optimization at runtime, because it has effectively solved the problem millions of times already.

The comparison shows that learning did not make locomotion simpler. It changed which problem engineers solve, moving the difficulty from runtime robustness into simulation design and reward engineering.

How Do Robots Learn to Move Like Humans?

Early reinforcement learning policies walked reliably but did not look human. No human data was involved at all. Instead, reward functions consisted of ten to fifteen hand-written terms. Keep the upper body upright, lift the feet high enough, stay near a default pose, follow a commanded velocity.

Two research directions changed that.

Motion Imitation and Retargeting

The first approach records human movement with motion capture, then trains the policy to reproduce it. This replaces a stack of hand-tuned reward terms with a single objective. As Frank summarizes: "it boils down to, yeah, track what the human is doing."

One step makes this possible. A human skeleton and a robot have different proportions and different joint structures, so the recorded motion cannot be applied directly. Retargeting maps the human trajectory onto a motion the robot can physically execute within its own kinematic constraints. The policy then simply tracks that retargeted trajectory.

This line of work traces back to DeepMimic, published in 2018, which introduced reward functions based purely on tracking a reference motion.

Adversarial Motion Priors

The second approach does not track one specific motion. It learns what human movement looks like in general.

A discriminator network is trained to judge whether a motion came from a human or a robot. The policy then earns reward for fooling it. The technique was introduced as Adversarial Motion Priors in 2021. A robot learns to walk forward and stay upright while moving in a recognizably human way. Nobody has to define that style in equations.

What Does This Mean for Manufacturing?

Legged robots themselves are not what factories need most. A wheeled platform is more stable and more efficient for moving material across a flat plant floor. The valuable output of legged robotics research is the method, not the form factor.

That method is now moving into manipulation, which is exactly where Frank's work sits today. The hard problem in a factory is not walking. It is contact: gripping a part whose position varies, applying the right force, handling objects that are not identical every time. The classical stack struggled with additional contacts for the same reason it struggled with opening a door.

This is the foundation of Physical AI, and it is what RobCo builds on with AI in robotics applied to real production environments:

• Learned behavior where variation is the problem: Autonomous Industrial Robots handle applications such as machine tending, palletizing, and materials handling. They also cover bin picking and sorting in unstructured environments, where every part is presented differently.
• Simulation before deployment: The same digital twin and domain randomization pipeline that produced walking policies is what makes a robot robust to real-line variation.
• Modular hardware, reconfigured not replaced: The modular robots are built from interchangeable components. A new task means swapping modules, not replacing the whole robot. Development happens at the Munich headquarters, with manufacturing in Germany and the US (Austin, TX).
• Operable by the people on the floor: Programming runs through no-code software (RobFlow). Shop floor employees set up and adjust the robots without a robotics background, automating the ordinary so people can focus on higher-value work.

For a closer look at how the layers fit together, from instruction to physical motion, our article How autonomous robots work breaks down the full stack.

Want to know which tasks in your production could be automated today? Book a consultation.

Frequently Asked Questions About Legged Robots

What are legged robots?

Legged robots are machines that move on legs rather than wheels or tracks, including quadrupeds and humanoids. They renegotiate ground contact with every step, which makes balance and locomotion far harder to control than wheeled movement across flat, predictable surfaces.

How do legged robots learn to walk?

Most legged robots today learn to walk through reinforcement learning in simulation. A policy is trained by trial and error across millions of simulated runs, guided by reward functions or by tracking recorded human motion, then transferred to real hardware.

What is domain randomization?

Domain randomization is a training technique that deliberately varies simulated physics, including friction, mass distribution, and external pushes. Forcing a policy to succeed across all these variations produces behavior robust enough to run on real hardware it has never encountered before.

What is sim-to-real transfer?

Sim-to-real transfer means deploying a policy trained entirely in simulation onto a physical robot without additional real-world training. It works because Newtonian physics simulates accurately and because randomized training covers the gap between simulation and reality.

Are legged robots useful in factories?

Legged robots are rarely the right form factor for factory work, since wheeled platforms are more stable and efficient on flat floors. The valuable result is the method: the simulation and learning pipeline developed for locomotion now enables robots to handle unstructured manipulation tasks.

‍