Robot Locomotion Explained: How a Robot Picks a Footstep

Walking looks continuous and it isn’t. A legged robot makes a discrete choice several times a second: out of all the places that foot could land, which one? Robot locomotion explained properly starts there, because footstep selection is where geometry, friction and momentum all have to agree at once. Get it wrong by three centimeters and the robot either trips or sails past its own center of mass and goes down. A human toddler solves this in about eighteen months of practice. Engineers have been working on it since the 1970s and it’s still where legged robots fail.

Key Takeaways

• Every candidate footstep has to pass three tests simultaneously: kinematic reachability, contact validity, and dynamic correctness. Failing any one of them drops the robot.
• The dynamics test is the counterintuitive one. A robot doesn’t step where it wants to go, it steps where the step will cancel the momentum it already has.
• Most legged robots don’t carry a 3D model of the world. They carry a 2.5D elevation map: one height value per grid cell. It’s cheap and it works for stairs, rubble and hills.
• Learned controllers now outperform hand-tuned ones on rough ground. Miki and colleagues, in Science Robotics (January 2022), trained a quadruped whose controller completed an alpine hiking route inside the posted human time.
• The honest failure mode is perception, not control. The same paper notes depth sensing degrades with “difficult lighting, dust, fog, reflective or transparent surfaces,” and the fix was teaching the robot to trust its own joints when its eyes lie.

Test one: can the leg actually get there?

Kinematic reachability is the question of whether a joint configuration exists that puts the foot at a given point, without exceeding any joint limit or colliding with the robot’s own body.

This sounds trivial and isn’t. A leg has a workspace shaped roughly like a curved shell, not a sphere, and that shell moves with the body. A spot 40 centimeters ahead might be reachable when the hips are low and unreachable a tenth of a second later when they’ve risen. So reachability is evaluated against where the body will be at touchdown, not where it is now. The robot is solving for a target in the future.

There’s also a self-collision constraint that produces the odd high-stepping gait you see in robots crossing debris: the swing leg has to travel a path that doesn’t clip the stance leg, the body, or the obstacle it’s stepping over. That’s a trajectory problem, not just an endpoint problem.

Test two: will the ground hold?

Contact validity asks whether the surface at that point can supply the force the robot needs without slipping or breaking.

Three things get checked. Is it solid? Loose gravel and thin ice look the same to a depth camera. Is it big enough? A foot needs a patch, and the edge of a stair tread is a terrible place to land because half the sole hangs over nothing. Is the friction sufficient? The force the robot wants to apply has to stay inside what engineers call the friction cone: push too far sideways relative to downward, and the foot slides.

The sensing problem here is real. Thin, shiny, transparent and dusty surfaces defeat depth sensors, and that’s exactly the list in Miki et al.’s 2022 Science Robotics paper, which is why their controller used an attention-based recurrent encoder to blend camera data with proprioception: the robot’s sense of its own joint positions and forces. The learned behavior is the interesting part: the controller learned when to stop believing the camera and walk by feel. Anyone who has walked down an unlit staircase has run the same policy.

Test three: does the step fix the momentum?

This is the test that surprises people, and it’s the heart of robot locomotion.

Picture a human leaning forward until they have to catch themselves. Where do they put the foot? Not under their body: out in front, past the lean, at the place where planting it will arrest the fall. If they place it under their center of mass, they keep falling. The forward step converts horizontal momentum into a rotation about the new contact, which bleeds the momentum off.

Engineers formalize this with simplified models. Treat the robot as a point mass on a massless leg (the linear inverted pendulum) and you can compute, in closed form, the point where a footstep will bring the system to rest. Variants of that quantity get called the capture point or the divergent component of motion. A whole-body controller then figures out all the joint torques needed to get the foot there while keeping the torso upright and the arms out of the way.

So the actual decision loop is: estimate my current momentum, compute where a foot would have to land to neutralize it, check whether that place is reachable and solid, and if not, pick the best compromise and plan to recover on the step after. Several times a second. Forever.

How to Teach Your Kid About Robot Locomotion

Every one of these concepts is available on a sidewalk.

Ages 5–8: Stepping stones, then the lean

Chalk a line of circles with uneven spacing. Have your kid cross without missing. Then move one circle just out of easy reach and watch what their body does: they’ll crouch, swing an arm, or take a smaller prior step. Name it: “you changed the step before the hard one.” That’s footstep planning over a horizon, and they just did it without being told.

Then the lean: stand with feet together, lean forward slowly until a foot shoots out. Ask whether it landed under them or in front. Always in front. Ask why.

Ages 9–12: Friction cone with a shoe

Put a shoe on a smooth board and tilt the board slowly until the shoe slides. Mark the angle. Repeat with a different shoe, a wet board, a towel. They’ve just measured a friction limit, which is the thing deciding whether a robot’s foot holds. Then ask the real engineering question: how would a robot know the angle before stepping? It mostly can’t, which is why robots step conservatively.

Ages 13+: Build an elevation map by hand

Draw a 10-by-10 grid on paper. Have your kid walk around a cluttered room measuring the height of whatever is at each grid cell (floor, chair leg, book stack) and write one number per cell. Then ask them to plan a route for a robot with a maximum step height of 15 centimeters. Then the kicker: ask them to represent a table. One height value per cell cannot describe “there’s a surface at 75 centimeters and empty space underneath.” That limitation is exactly why a humanoid on monkey bars is hard, which we cover in humanoid balance control on the monkey bars.

The question to ask: “If you’re already falling forward, should you step in front of yourself or underneath yourself? Why?”

Model-based and learned footstep planning, compared

	Model-based planning	Learned policy (reinforcement learning)
How the step is chosen	Solve a simplified dynamics model online	A neural network maps sensor history to joint targets
Where the knowledge lives	In explicit equations an engineer wrote	In weights trained over millions of simulated steps
Strength	Predictable, analyzable, provable in limited cases	Handles terrain variety and sensor noise far better
Weakness	Brittle when the model’s assumptions break	Hard to verify; can fail in ways nobody anticipated
Perception needs	A clean elevation map	Learns to weigh unreliable perception against proprioception
Demonstrated result	Decades of reliable flat-ground and stair walking	Alpine hike inside human time (Miki et al., 2022)
Agility result	Conservative gaits	2 m/s across consecutive obstacles (ANYmal Parkour, 2023)
Certification status	Easier to argue to a safety authority	An open problem

That last row matters more than it looks. A learned controller that works beautifully is still a function nobody can fully explain, and safety authorities have no established framework for approving one. IEEE Spectrum’s September 2025 reporting noted that new ISO standards for dynamically balancing legged robots were still under development, with Boston Dynamics’ Matt Powers saying the company would “start with relatively low-risk deployments, and expand as we build confidence.”

What this changes about how you read robot videos

Watch the feet, not the body

In any legged-robot video, the informative thing is foot placement. Are the steps evenly spaced on flat ground, which proves nothing? Or are they adjusting: short step, long step, a foot placed on the only solid-looking patch? Adjusted spacing on varied terrain is evidence of real perception in the loop.

Flat floors hide everything

A robot walking across a trade-show carpet is demonstrating almost nothing about locomotion, because the three tests are trivially satisfied everywhere. The hard demos involve gravel, stairs with inconsistent risers, grass hiding holes, and wet tile.

A stumble that recovers is better than a clean walk

Recovery is the capability. A robot that gets shoved, takes a wide corrective step and keeps going has demonstrated that its momentum computation and its planner are both working under disturbance. A robot that never gets perturbed has demonstrated that the demo was controlled.

The arms are part of the legs

Watch a humanoid’s arms during a recovery step. They swing, and not for show: swinging an arm changes the body’s angular momentum without needing a new contact, which buys time for the leg. Digit 5, Agility’s industrial humanoid covered by IEEE Spectrum in September 2026, takes that logic to its conclusion: if a person approaches, it sets its load down and sits, because a seated robot can’t fall on anyone. Momentum management all the way down. Our comparison of industrial versus home humanoids covers that design choice.

What not to do

Don’t explain this to a kid as “the robot calculates physics.” That’s true and useless. Explain it as three questions the robot asks about every possible place to put its foot, and have them ask the same three questions while crossing a creek on rocks. The concept survives; the jargon doesn’t need to.

What to Watch For Over the Next 3 Months

• Week 4: Look for a published success rate on any terrain-crossing robot result, attempts versus completions. Papers usually report it; product videos never do. The gap between those two cultures is the most useful media-literacy lesson in robotics.
• Month 2 red flags: Locomotion demos filmed exclusively indoors on uniform flooring, or edited so you never see a full gait cycle. Also any claim of “human-level walking” without a terrain specification, which is a meaningless phrase on its own.
• Month 3 self-check: Can your kid explain why a robot steps in front of itself when it’s falling forward? If they can, they understand the single non-obvious idea in legged locomotion, and it will stick for years.

Frequently Asked Questions

Why do robots walk with bent knees?

A straight leg is a kinematic singularity: at full extension the leg can’t adjust length, so it loses the ability to absorb or apply vertical force smoothly. Keeping a bend preserves control authority in both directions. It costs energy, which is why it looks awkward and why better designs are a continuing research topic.

Do robots plan more than one step ahead?

Yes, and that’s necessary. Hoeller and colleagues’ ANYmal Parkour work (2023) used a high-level navigation policy selecting among hierarchical skills (walking, jumping, climbing, crouching), which is planning over a horizon rather than step to step. Horizon planning is what lets a robot shorten a step now to set up a long one next.

Is walking harder than running for a robot?

Slow walking is harder than it looks, because during the slow phases the robot has little momentum to work with and depends more on precise static balance. Running is harder in a different way: everything happens faster and there are flight phases with no contact at all. Neither is “easy.”

Why don’t they just use wheels?

Often they should, and the best engineers say so. Wheels are far more efficient and cannot fall over. Legs only win where the terrain has steps, gaps or debris: stairs, rubble, forests, construction sites. Hello Robot’s assistive Stretch robot is wheeled on purpose, which we cover in assistive robots as robotics’ quiet branch.

What’s the single biggest unsolved part?

Perception reliability in bad conditions, by most accounts. The control math is well developed. Knowing whether the thing in front of you is a solid step or a shadow on a puddle is not, and it’s where learned controllers currently earn their keep by learning to distrust their own cameras.

Can a kid simulate this at home?

Yes, with free tools. Simple physics sandboxes and open-source robotics simulators run on an ordinary laptop, and the classic first exercise, get a two-link leg to balance a point mass, teaches the capture-point idea directly. Start small: a pendulum before a biped.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Miki, T., Lee, J., Hwangbo, J., Wellhausen, L., Koltun, V., & Hutter, M. (2022). “Learning robust perceptive locomotion for quadrupedal robots in the wild.” Science Robotics, 7(62). https://arxiv.org/abs/2201.08117
2. Hoeller, D., Rudin, N., Sako, D., & Hutter, M. (2023). “ANYmal Parkour: Learning Agile Navigation for Quadrupedal Robots.” https://arxiv.org/abs/2306.14874
3. IEEE Spectrum. (2026, September 11). “Video Friday: Humanoid Robot Takes On Monkey Bars.” https://spectrum.ieee.org/video-friday-disaster-response-robots
4. Ackerman, E. (2025, September 11). “Reality Is Ruining the Humanoid Robot Hype.” IEEE Spectrum. https://spectrum.ieee.org/humanoid-robot-scaling
5. Ackerman, E. (2026, September 15). “Digit 5 May Be the First Humanoid Robot Worker That’s Truly Safe.” IEEE Spectrum. https://spectrum.ieee.org/humanoid-robot-safety
6. Occupational Safety and Health Administration. “Industrial Robot Systems and Industrial Robot System Safety.” OSHA Technical Manual, Section 4, Chapter 4. https://www.osha.gov/otm/section-4-safety-hazards/chapter-4
7. International Federation of Robotics. (2026, September 24). “Five Million Robots now Operate in Factories Globally.” https://ifr.org/ifr-press-releases/news/five-million-robots-now-operate-in-factories-globally