How humanoid robots keep their balance, explained without equations

A humanoid robot keeps its balance by measuring the state of its own body many times a second and correcting it continuously. Sensors report joint angles and the tilt and acceleration of the torso, and a controller chooses how hard each motor should push. Today that controller is often a neural network trained by trial and error in simulation. It can absorb ordinary bumps, but not every push or floor.

Why walking on two legs is hard

A standing person or robot is balanced when its centre of mass sits above the area covered by its feet. Walking breaks this rule on purpose: with each step the body tips forward and is caught by the next foot. A two-legged robot is therefore always falling a little and recovering, which is why IEEE Spectrum's robotics editor describes bipedal robots as fundamentally unstable in a September 2026 article on Agility Robotics' Digit 5.

The practical consequence is simple. A wheeled cart that loses power stops. A walking humanoid that loses power, or whose controller makes a bad decision, can fall over. Everything below is about how robots reduce the chance of that happening.

What the robot senses

Most balance control relies on what engineers call proprioception: the robot's sense of its own body, rather than of the world around it. Two kinds of sensor do most of the work.

• Joint encoders report the angle and speed of every joint. Unitree, for example, lists a dual encoder per joint on its G1 humanoid.
• An inertial measurement unit (IMU) measures how the body is tilting, turning and accelerating, in a similar role to the human inner ear.

In a widely cited UC Berkeley study, the walking controller for Agility Robotics' Digit received only joint encoder and IMU data, with no cameras at all, according to the paper on real-world humanoid locomotion with reinforcement learning.

Cameras and lidar matter for seeing steps, obstacles and people ahead. Unitree lists a depth camera and a 3D lidar on the G1. But the fast, reflexive part of staying upright usually comes from the body's own signals, because they arrive with little delay and do not depend on lighting or a clear view.

From hand-written rules to learned controllers

For many years, walking robots were controlled by carefully engineered models: equations describing the robot's body and the ground, used to plan where each foot should go. These methods work well when reality matches the model. Boston Dynamics explains in its post on adding reinforcement learning to Spot that real sites contain conditions that are hard to observe and model, such as a floor that may be wet or greasy.

Reinforcement learning takes a different route. Instead of programming every rule, engineers build a simulated robot, give it a goal such as walking at a requested speed without falling, and let a neural network try millions of times. Behaviour that works is rewarded; behaviour that ends on the floor is not. Over time the network learns which motor commands keep the body upright.

The key trick is variety. In the Berkeley Digit study, the researchers randomised the robot's physical properties, its control settings and the simulated physics, and added noise and delay to the sensor readings. A controller that copes with all those variations in simulation is more likely to cope with the one real robot and the one real floor it eventually meets.

How the learned controller behaves

The Berkeley controller is a type of neural network called a transformer, which looks at a recent history of sensor readings and its own actions to decide the next move. The authors report that this history lets it adapt its behaviour on the fly without retraining. It ran at 50 decisions per second, while lower-level motor control ran at 1,000 times per second.

Transferred straight from simulation to hardware, it walked Digit across plazas, sidewalks, running tracks and grass. The researchers tested it by throwing a large yoga ball at the robot, pushing it with a wooden stick and pulling it from behind.

A follow-up study by members of the same group, submitted in October 2024, reports that a similar camera-free approach carried a humanoid over more than four miles of hiking trails in Berkeley and up steep San Francisco streets, according to the paper on humanoid locomotion over challenging terrain.

The idea is not limited to humanoids. A 2020 Science Robotics paper from ETH Zurich showed a four-legged ANYmal robot, trained in simulation and using only body sensing, crossing mud, snow, rubble and vegetation, as described in Learning quadrupedal locomotion over challenging terrain.

Whole-body control: balance while doing something else

Walking in a straight line is only part of the job. A robot that waves, carries a tray or kicks a ball moves its centre of mass with every arm and leg movement, and the controller has to compensate. Coordinating all joints together towards one goal is usually called whole-body control.

Boston Dynamics describes this in a post on teaching Atlas a football move. Engineers recorded a footballer with motion capture, adapted the movement to Atlas's different body, then trained it with reinforcement learning across thousands of parallel simulations. The company says this gave the equivalent of a year of trial and error in 24 hours, and that for almost all of Atlas's skills the learned policy worked on the real robot at the first attempt. It frames football as useful because it demands moving and manipulating at the same time.

This is also why choreographed dance routines and gestures need preparation time. Each new movement changes how the robot's weight shifts and has to be developed and tested for balance.

Hybrid systems in commercial robots

Commercial robots do not always replace engineered control entirely. Boston Dynamics says Spot's locomotion combines a model predictive controller with a learned policy trained in more than a million simulations. Before release it was checked over thousands of hours of simulated and real-world running, on an internal fleet logging more than 2,000 hours a week. The company reports the update reduced falls on slippery and irregular surfaces and cites a mean time between falls of hundreds of hours on that fleet.

That figure is a useful reminder: even a mature commercial legged robot is described in terms of how rarely it falls, not as a robot that never falls.

Why pushes and uneven floors still cause falls

Learned controllers are robust, but they have limits, and the published research is open about them.

• Disturbances beyond what was trained. The Berkeley authors note that excessive external disturbances, such as a very strong pull on a cable, can still defeat the controller. A robot can only recover from a push if it can place a foot fast enough and far enough to catch itself.
• The gap between simulation and reality. Simulated physics never matches reality exactly. The same paper reports imperfect speed tracking and motors that responded asymmetrically on the real robot.
• Surfaces the robot cannot feel in advance. A controller that relies on body sensing learns about a slippery patch, loose cable cover or soft carpet edge only when a foot is already on it. Cameras help, but only if the perception system recognises the hazard.
• Busy hands and fast moves. When the arms are carrying, gesturing or performing, there is less margin left for correcting a stumble.
• Hardware and power. Overheating motors, low batteries, a lost connection or a software fault can all cut the control loop that keeps a walking robot upright.

Researchers are also working on what to do when balance is lost. IEEE Spectrum reported in 2022 that a team at Inria trained the TALOS humanoid to reach out and brace against a nearby wall, and that it avoided falling in three of four trials where it otherwise would have fallen. That was a laboratory result on a stationary robot, not a feature buyers can expect today.

Agility Robotics has taken a different, preventive route with Digit 5. According to IEEE Spectrum, when people come close the robot is designed to avoid them, stop, or put down its load and lower itself into a seated position, so that it cannot fall onto anyone. The magazine describes the approach as not elegant, but effective. Work on a dedicated safety standard for legged robots is covered in our article on ISO 25785-1.

What this means for event organisers

This article is general information, not legal or safety-engineering advice; your supplier and venue remain responsible for their own risk assessments.

• Treat the floor as part of the act. Smooth, level, dry surfaces without loose cables or ramp edges give any robot more margin. Our venue checklist covers what to ask.
• Keep a clear zone around walking robots. A robot that is pushed or bumped by an excited guest may fall, and a 35 kg robot such as the Unitree G1, by its maker's figure, is heavy enough to hurt someone. Plan barriers and staffing as described in our public events safety guide.
• Ask how new moves were tested. Custom gestures, dances or carrying tasks change the robot's balance. Ask the supplier whether each routine has been rehearsed on similar flooring.
• Plan for a fall or a pause. Agree in advance who stops the robot, how it is moved, and what the show does if it has to sit out. See what happens if the robot fails on the day.

Frequently asked questions

Do humanoid robots use cameras to keep their balance?

Mostly not for the fast part. Balance usually depends on joint encoders and an inertial measurement unit, which report the robot's posture and movement with little delay. A UC Berkeley study controlled Agility's Digit using only those signals, without cameras. Cameras and lidar are still useful for spotting steps, obstacles and people ahead, so many robots, such as the Unitree G1, carry a depth camera and lidar as well.

What does reinforcement learning mean for a walking robot?

It means the walking controller is learned by trial and error rather than written by hand. Engineers simulate the robot, reward it for walking without falling, and let a neural network practise millions of times under varied conditions such as different friction, weights and sensor delays. The trained network is then transferred to the real robot. Boston Dynamics, Unitree and university research groups all describe using this approach.

Can a humanoid robot recover if someone pushes it?

Often, within limits. Researchers have tested learned controllers by pushing robots with sticks, throwing balls at them and pulling them from behind, and the robots stayed upright. The same researchers note that excessive disturbances can still cause falls. Recovery depends on how hard the push is, what the robot is doing at that moment and the surface underfoot, so crowds should not be allowed to push or lean on robots.

Why do robots fall on surfaces that look flat?

A floor that looks flat to people can still be slippery, soft, slightly sloped or interrupted by cable covers and carpet edges. A controller that relies on body sensing only detects these conditions once a foot is on them. Simulation never matches every real surface exactly, so unfamiliar flooring is a common source of stumbles. Testing on the actual venue floor before the event reduces the risk.

How this article was made: written by the humanoid4rent editorial team with AI assistance; every factual statement was checked against the sources listed below on 2026-10-31. Spotted an error? Write to info@humanoid4rent.com.