Technology
How Humanoid Robots Actually Walk: The Engineering Inside
Bipedal locomotion is the foundational capability of humanoid robotics. This deep-dive covers the control theory, sensor hardware, actuator design, and reinforcement learning techniques that make modern robots walk, along with a comparison of approaches from Figure, Agility, Boston Dynamics, and Unitree.
A humanoid robot that can't walk reliably can't do anything useful. Before it picks up a part, navigates a warehouse aisle, or climbs a staircase, it has to solve a problem that took biological evolution roughly 4 million years to refine: staying upright on two legs while moving through an unpredictable world. Bipedal locomotion is the foundational capability that separates a humanoid robot from a wheeled machine, and it remains one of the hardest unsolved problems in robotics.
The difficulty isn't speed or strength. It's stability. A two-legged robot is inherently unstable — like an inverted pendulum constantly on the verge of toppling. Walking requires predicting ground contact forces, managing center of mass through every step, adapting to uneven surfaces, and recovering from unexpected pushes, all in real time at 1 kHz control loop frequencies. The gap between what Boston Dynamics' Atlas demonstrated in viral videos and what can be deployed reliably in real workplaces is measured in millions of engineering hours.
Key Stats
1 kHz
Joint-level control frequency
~40ms
Typical human step recovery window
28-65+
Degrees of freedom (by model)
6-axis
Force/torque sensors per foot
How Bipedal Walking Actually Works
Human walking is what engineers call a "controlled fall." Each step begins with the body leaning forward, shifting weight onto the front foot while the trailing leg swings through. The center of mass (CoM) actually traces an arc that dips during single-leg support (when one foot is the only contact) and rises slightly as both feet briefly contact the ground. It's an energy-efficient oscillating motion that recovers kinetic and potential energy through each stride.
For robots, reproducing this involves several interrelated control problems that must be solved simultaneously:
• Zero-Moment Point (ZMP) control: The ZMP is the point on the ground where the net moment of all ground reaction forces is zero. If the ZMP stays within the convex hull of the support polygon (the foot contact area during single support, or the region between both feet during double support), the robot won't topple. Early humanoids like Honda's ASIMO used ZMP-based controllers exclusively, which made them stable but slow and shuffling because they kept both feet close to the ground to maintain a wide support polygon.
• Capture Point control: A refinement of ZMP theory that predicts where the robot must step to recover from a perturbation. The "capture point" is the foot placement location that would bring the robot to a complete stop without falling. By planning steps that continuously track the capture point, a controller can produce more natural, dynamic gaits that aren't limited by keeping the ZMP strictly inside a static footprint.
• Whole-Body Control (WBC): A framework that treats all the robot's joints as a unified system, allocating torques across every joint to satisfy multiple objectives simultaneously — maintain balance, track a desired trajectory, avoid joint limits, minimize energy — all ranked by priority. WBC runs at high frequency (typically 200-500 Hz) and is computationally intensive but produces natural-looking, coordinated motion.
• Model Predictive Control (MPC): Plans a sequence of future footsteps and body motions over a short prediction horizon (typically 0.5 to 2 seconds), optimizing for a cost function that includes stability, energy, and task completion. MPC can plan around upcoming terrain features and is increasingly used in conjunction with WBC for the full locomotion stack.
The shift from purely model-based control to hybrid model-plus-learning approaches has been one of the most important developments in legged robotics since 2020. Rather than relying entirely on physics models that are difficult to tune for every terrain type, controllers now include components trained by reinforcement learning (RL) in simulation. The simulation runs thousands of times faster than real time, exposing a virtual robot to millions of fall scenarios, terrain types, and perturbations that would take years to collect in the real world.
Under the Hood: Sensors, Actuators, and the Compute Stack
Walking well requires knowing, at every millisecond, exactly where the robot's body is, what forces its feet are experiencing, and what the terrain ahead looks like. This requires a dense sensor suite working in tight coordination with the actuators and the onboard compute.
AI-generated image
Articulated knee and ankle joints are the most mechanically stressed components in bipedal locomotion, experiencing impact loads several times the robot's body weight during each heel strike.
The Sensor Stack
• Inertial Measurement Unit (IMU): A 6-axis sensor (3-axis accelerometer plus 3-axis gyroscope) that measures body orientation and angular velocity at rates up to 1 kHz. The IMU is the primary source of real-time body state estimation and is typically located in the robot's pelvis to minimize noise from limb motion.
• Joint encoders: High-resolution rotary encoders on each joint measure angular position. Combined with current sensors in the motor drives, they give the controller exact knowledge of joint angles and estimated torques. Resolution of 18 to 20 bits (better than 0.001 degrees) is typical for precision walking robots.
• Foot force/torque sensors: Six-axis F/T sensors embedded in each foot measure ground contact forces in all directions. These sensors detect when a foot touches down, how hard, and whether the contact is stable or shifting. They're essential for the reactive balance corrections that keep the robot from falling when it steps on an unexpected surface.
• Depth cameras and LiDAR: Forward-facing depth sensors (structured light cameras, time-of-flight cameras, or LiDAR) build a real-time 3D map of the terrain ahead. This terrain map is fed into the footstep planner, which selects safe foot placement locations and adjusts step height and timing for obstacles, stairs, and uneven ground.
• Proprioceptive actuator feedback: Modern torque-controlled actuators can measure the force they're exerting directly, creating a closed-loop force controller at the joint level. This allows the robot to detect unexpected resistance (stepping on a slippery surface, a foot catching on an edge) and respond within milliseconds.
Actuator Design Choices
The transition from hydraulic to electric actuators has been the most important hardware shift in humanoid robotics in the past decade. Boston Dynamics' original Atlas used hydraulic actuation — powerful and back-drivable, but heavy, complex, and requiring a hydraulic power unit. The new electric Atlas, revealed in 2024, shifts to custom rotary and linear electric actuators throughout.
⚡ Series Elastic Actuators (SEA)
A spring element in series with the motor allows force to be estimated from spring deflection, provides impact tolerance, and creates inherently safe physical interaction. SEAs were pioneered at MIT and used extensively in NASA's Valkyrie robot. Downside: reduced bandwidth compared to rigid actuators.
🔄 Quasi-Direct Drive (QDD)
A low-ratio gearbox (typically 6:1 to 9:1) between the motor and output link. High bandwidth, back-drivable, and capable of measuring output torque from motor current. Used in MIT Cheetah robots and influenced many current humanoid designs. Requires larger, more powerful motors than high-ratio geared designs.
⚙️ High-Ratio Geared Electric
Traditional approach using high gear ratios (100:1 or more) for high torque from smaller motors. Not back-drivable — meaning the robot can't be pushed through its joints easily — which creates safety concerns and makes force control difficult without additional sensors. Used in many industrial robots and some early humanoids.
🌊 Linear Hydraulic (Legacy)
High power density, excellent back-driveability, and the ability to generate very high forces in compact actuators. The original Atlas and PETMAN used hydraulics. Still used in some applications, but operational complexity, leak risk, and efficiency losses make it unfavorable for commercial deployments.
| Robot | Height / Weight | Actuator Type | DOF | Walking Speed | Status (2026) |
|---|---|---|---|---|---|
| Boston Dynamics Atlas (Electric) | 1.50 m / 89 kg | Electric rotary + linear | 28 | ~1.5 m/s | Commercial (Hyundai trials) |
| Figure 02 | 1.70 m / 60 kg | Electric QDD | 65+ | ~1.2 m/s | Deployed (BMW plant) |
| Tesla Optimus Gen 2 | 1.73 m / 57 kg | Electric (custom) | 28 | ~0.5 m/s (reported) | Internal (Tesla factories) |
| Agility Robotics Digit | 1.75 m / 65 kg | Electric SEA + QDD | 20 | ~1.5 m/s | Deployed (Amazon warehouses) |
| Unitree H1 | 1.80 m / 47 kg | Electric QDD | 19 | 3.3 m/s (record) | Available commercially |
| Fourier GR-1 | 1.65 m / 55 kg | Electric SEA | 40 | ~1.0 m/s | Limited commercial |
Who's Building the Walking Robots
The current generation of bipedal robots draws on decades of academic research from MIT, Carnegie Mellon, Georgia Tech, and ETH Zurich, now being commercialized by a cluster of well-funded startups and established robotics companies.
• Boston Dynamics: The benchmark for athletic robot locomotion. The electric Atlas (2024) replaced the hydraulic research platform with a commercial-grade machine. Owned by Hyundai since 2021, Boston Dynamics is running trials in Hyundai's automotive plants. Its walking and dynamic movement capabilities remain industry-leading, though commercial deployment is newer than rivals like Agility.
• Agility Robotics: Oregon-based startup that has arguably moved fastest from lab to deployment. Digit is operating in Amazon warehouses doing tote transport. Agility raised a $150 million Series B led by Amazon's Industrial Innovation Fund in 2023. The company's RoboFab manufacturing facility in Salem, Oregon is designed to produce 10,000 Digit units per year.
• Figure AI: 2022 startup that raised $675 million in a Series B at a $2.6 billion valuation, with investors including OpenAI, Microsoft, and Nvidia. Figure 02 is deployed at BMW's Spartanburg, South Carolina facility. Figure has integrated OpenAI's models into a natural language interface for task instruction.
• Unitree Robotics: Chinese manufacturer offering the H1 humanoid robot at roughly $90,000, dramatically below Western competitors. The H1 set a bipedal running speed record of 3.3 m/s (about 7.4 mph) in 2024. Unitree's price point is disruptive and has pushed competitors to accelerate cost reduction roadmaps.
• 1X Technologies: Norwegian startup backed by OpenAI, developing the Neo humanoid for household and commercial tasks. Unlike most competitors focused on factories, 1X explicitly targets less structured environments, which raises the locomotion difficulty considerably.
AI-generated image
Current deployments focus on structured warehouse and factory environments where terrain is predictable and safety protocols can be established — a constraint that reduces the locomotion difficulty compared to general-purpose outdoor use.
What Good Locomotion Unlocks for Industry
The practical case for bipedal robots in industry isn't purely about matching human mobility. It's about fitting into infrastructure designed for humans. Warehouses, factories, and logistics facilities have been built for workers who walk through aisles, climb stairs, step over floor cables, and navigate tight clearances between shelving. Wheeled robots require ramps, wide pathways, and floor markings. A robot that walks like a human can use the same infrastructure immediately.
Current deployments reveal both what's working and what's not. The Amazon-Agility partnership puts Digit doing tote transport in existing fulfillment centers — a highly repetitive task in a controlled environment where the robot walks predictable routes. This application works because the terrain is flat, clean, and consistent. Digit doesn't need to climb stairs or navigate a loading dock during its current deployment scope.
The BMW-Figure partnership goes a step further, involving part handling that requires the robot to approach workstations, grasp specific components, and place them precisely. Walking is still largely on flat factory floors, but the robot must stop, stabilize, and perform manipulation tasks — requiring the locomotion and manipulation systems to work together without the robot losing balance during a reach or a push.
The Stair-Climbing Reality Gap
Nearly every humanoid robot company shows video of their robot climbing stairs. Very few real deployments involve stairs at all. Stair climbing requires precise foot placement on step edges with limited margin for error, significant torque from the knee and hip joints as the robot lifts its full body weight per step, and reliable foot contact detection to avoid catastrophic mid-step failures. A fall on stairs is both more likely and more damaging than a fall on flat ground. Current commercial deployments are largely solving the "walk on flat floors reliably" problem first, with stair climbing still primarily a demonstration capability rather than a production deployment feature.
What's Coming Next in Bipedal Locomotion
The next 12 to 18 months of development are likely to be defined by three trends:
First, reinforcement learning in locomotion will move from the lab to production. Boston Dynamics and Agility have already deployed RL-trained components in their locomotion stacks for specific terrain types. The expansion of RL to cover a wider range of real-world conditions — different floor materials, gentle slopes, threshold crossings, surface variations — will improve reliability without manual tuning for every environment.
Second, energy efficiency will become a commercial differentiator. Most current humanoid robots have battery lives of 2 to 4 hours under working conditions. Doubling that to 4 to 8 hours requires either larger batteries (adding weight, which hurts locomotion) or more efficient locomotion control (capturing and reusing mechanical energy, reducing unnecessary joint movements, optimizing gait for efficiency rather than pure speed). Companies that solve the energy problem first will have a significant operational advantage in real-world deployments.
Third, terrain perception will get dramatically better. Current robots primarily handle terrain reactively — they detect a step edge or slope change as their foot makes contact and then adjust. Anticipatory locomotion — seeing an obstacle 3 meters ahead and pre-planning the foot placement sequence to navigate it smoothly — will make gaits less jerky and more reliable in messy environments. This requires better real-time terrain reconstruction from onboard sensors and faster integration between the perception system and the footstep planner.
Frequently Asked Questions
Why don't all robots just use wheels instead of legs?
Wheels are more energy-efficient and mechanically simpler on flat, smooth surfaces. The reason humanoid robots use legs is the same reason facilities are built with stairs, elevated platforms, uneven floors, and irregular terrain: the world was designed for walking humans. A wheeled robot requires infrastructure modifications — ramps, smooth pathways, cleared floor markings — that can cost more than the robot itself at scale. Legs allow the robot to use existing infrastructure without modification, which is the economic case for bipedal design in industrial settings.
How do robots handle unexpected pushes or slipping?
Disturbance recovery is handled by the same whole-body controller that manages normal walking, but with higher-priority balance objectives activated. When the IMU detects unexpected angular velocity (a shove) or foot force sensors detect a slip, the controller runs a recovery step: it computes a new capture point (where it must step to avoid falling) and places the foot there within 150 to 300 milliseconds. Boston Dynamics' Atlas has been demonstrated recovering from hockey stick shoves and unexpected pushes from multiple directions. Commercial robots have less athletic recovery but are tuned for the specific disturbances their deployment environment presents.
What is the fastest bipedal robot?
As of 2024-2025, Unitree's H1 holds the speed record for a bipedal humanoid at 3.3 meters per second (about 7.4 miles per hour). Boston Dynamics' Atlas can run at similar speeds in controlled demonstrations. For reference, an average human walks at 1.4 m/s and jogs at 2.5 to 3 m/s. However, top speed in a demo and reliable walking speed in a production environment are very different metrics. Most deployed commercial robots operate at 0.5 to 1.5 m/s for safety and reliability reasons.
How does reinforcement learning improve robot walking?
Reinforcement learning trains a control policy by simulating millions of walking scenarios and using a reward function to score each outcome (didn't fall = positive reward; fell = negative reward; efficient motion = bonus). The simulation runs thousands of times faster than real time, so a policy that would take years to develop through real-world trial and error can be trained in days or weeks. The resulting policy is then transferred to the real robot, often with a "sim-to-real gap" that requires additional fine-tuning. RL excels at producing controllers that generalize to terrain variations because the training distribution explicitly includes diverse surfaces, slopes, and disturbances.
Can current humanoid robots climb stairs reliably?
Boston Dynamics' Atlas and several research platforms can climb stairs reliably in controlled settings. Commercial deployments as of early 2026 largely avoid stairs due to reliability and safety concerns. Agility's Digit can climb stairs in demonstrations but Amazon's current deployments use only flat warehouse floors. Figure 02 has demonstrated stair climbing. The bottleneck is reliability, not capability — stairs present enough risk of falls that operators are conservative about deploying robots in stair-heavy environments until fall recovery and foot placement precision improve further.
The 12-Month Outlook
Bipedal locomotion is no longer the limiting constraint on humanoid robot deployments in controlled environments. Flat-floor walking is a solved problem at the level required for current applications. The active frontier is reliable locomotion in messier environments — light stairs, threshold crossings, loading docks, outdoor pavement — and energy efficiency that allows full shift operation without mid-shift recharging.
Watch for Agility's Amazon deployments to expand in scope and scale through 2026, providing the most detailed real-world data on commercial humanoid locomotion reliability. Unitree's aggressive pricing will push the entire sector toward cost reduction. And continued investment in simulation-based RL training will accelerate the terrain generalization that separates demonstration robots from ones that can be handed the keys to a facility and left to operate independently.
The Bottom Line: Bipedal locomotion in controlled industrial environments is production-ready. The competition has moved to reliability across diverse terrain, energy efficiency for multi-hour shifts, and the cost reductions needed to make humanoid labor economics work at scale.
The fundamental physics of walking hasn't changed. What has changed is our ability to build control systems sophisticated enough to handle it in real-time, at the cost and weight required for useful robots. The 4-million-year head start that evolution had is being closed, one simulation epoch at a time.