Robotics 4 November 2026 7 MIN READ
How Humanoid Robots Actually Work
Motors, sensors, balance and hands: what a humanoid robot is made of, what it does reliably today and why spec sheets flatter the machine.
THE SHORT ANSWER
A humanoid robot is an electric machine of motors, sensors and software that must balance continuously. Today's models lift tens of kilograms and work for hours in controlled sites. Hands, reliability and battery life are the limits, and vendor spec sheets show best cases rather than everyday performance.
On this page
From hydraulics to electric motors
For years the best-known acrobatic humanoid was hydraulic. Boston Dynamics retired that Atlas in April 2024 and unveiled a fully electric one aimed at commercial work. IEEE Spectrum’s analysis of the reveal explains the trade-off: hydraulics were long the only route to explosive, repeated motion such as jumps and flips, but dropping them removes hoses and widens the range of motion.
Researchers at ETH Zurich had summarised the same trade-off in 2019. Hydraulic legged robots store a lot of energy in a small space, but they are hard to scale down, noisy and smoky, so they mostly work outdoors. Electric actuators suit indoor work.
Today’s humanoids are electric machines with many joints. According to Boston Dynamics, the current electric Atlas stands 1.9 m tall, weighs 90 kg and has 56 degrees of freedom. Unitree lists its full-size H1 at about 180 cm and 47 kg, with a maximum joint torque of 360 N·m. Agility’s Digit 5 uses proprietary cycloidal actuators in its legs. These are all vendor figures, and we will come back to why that matters.
Seeing the world
A humanoid needs to know where its body is and what is around it. Agility says Digit relies on onboard cameras, depth sensors and inertial measurement units (IMUs), processed locally on the robot. Unitree’s H1 combines 3D lidar with a depth camera for 360° sensing. Atlas adds tactile sensing in its fingers and palm, plus a 360° camera view.
Staying upright
Balance is not a trick done once; it is a control problem solved continuously, because a two-legged body is always close to falling. Boston Dynamics says its robots combine classical simulation and model predictive control with learned reinforcement-learning and vision tools. In academic work, a 2025 paper showed whole-body model predictive control running in real time on both quadruped and full-size humanoid hardware.
Falling is also a safety problem, not just a technical one. Agility says Digit 5 is designed to respond to nearby people by stopping, avoiding them or sitting down, so it cannot fall on anyone.
Why hands are the hard part
If legs are mostly solved for flat floors, hands are not. Boston Dynamics calls hand design a “ruthless tradeoff” between dexterity, strength, ruggedness, cost, repairability and sensing. Its newest Atlas hand went from 7 to 13 degrees of freedom. At the cheap end, Unitree’s optional three-fingered hand for the G1 EDU has just 7 active degrees of freedom in total.
Tesla’s chief executive was blunt about this on the company’s third-quarter 2026 earnings call. Fingers must be soft enough for delicate objects yet survive constant wear, and a robot cannot heal like skin. He said Tesla is ramping up production before the hand is perfected: “We’re building robots that do not have perfected hands”, The Verge reported.
Reading a spec sheet
Runtime, payload and speed are the numbers everyone quotes, and they are the least comparable. Every figure below is published by the vendor and measured under a workload it does not define.
- Atlas (Boston Dynamics): 4 hours of battery, or 2 hours with heavy lifting, and a 3-minute automatic battery swap. Payload is 50 kg for an instant, 30 kg sustained and 20 kg with one hand.
- Digit 5 (Agility): a 90-minute battery that charges in 9 minutes, and a 22.7 kg (50 lb) lift. IEEE Spectrum reported that Agility’s former product chief explained that 60 of Digit’s 90 minutes are held as reserve, so in practice it works for about 30 minutes before it charges.
- G1 (Unitree): about 2 hours of battery, from US$13,500, with an arm that carries about 2 kg, which Unitree says varies greatly with arm posture.
Read the second line again. A headline of 90 minutes can mean 30 minutes of work. The headline is the best case; the working number is the one to ask for.
How a robot learns to walk
Most modern walking controllers are not written by hand. They are learned in simulation and then transferred to the real robot, which is known as sim-to-real.
The landmark paper came from ETH Zurich in 2019. Before it, reinforcement learning for legged robots usually needed weeks or months of real robot time, and the robot could behave chaotically while it learned. The researchers trained a neural-network policy for the ANYmal quadruped entirely in a simulator that ran about 1,000 times faster than real time, at nearly 500,000 steps per second. Each training session ran on a single desktop computer and took under 11 hours. A small neural network learned from the real robot (an “actuator net”) stood in for the joint dynamics that are hard to simulate. The learned controller then beat the best earlier ANYmal controller and learned to right itself after falls.
The next steps were about toughness and speed. In 2020, ETH and Intel trained a controller that used only the robot’s own sensors and transferred without further tuning to mud, snow, rubble, vegetation and water. It needed a two-stage method: a “teacher” that sees the true terrain, possible only in simulation, trains a “student” that sees what the real robot sees. By 2021, simulating thousands of robots in parallel on one graphics card cut training to under four minutes on flat ground and about twenty on uneven terrain.
These results are for specific robots. They show that walking can be learned cheaply and fast; they do not show that general-purpose manipulation is solved.
Why this is old news and new news at once
Robotics researchers have long noted that the “hard” problems for people, such as chess or tests of logic, are easier for machines than the “easy” ones, such as seeing, grasping and moving through a cluttered room. The roboticist Hans Moravec argued in the 1980s that perception and movement carry about a billion years of evolutionary refinement, while abstract thought is a very recent trick. That idea is still a useful way to read the humanoid boom: the legs and the learning have moved quickly, and the hands and the reliability are where the time goes.
What they do reliably today
Evidence from the field is narrow and comes mostly from companies. Agility reports that its previous-generation Digit moved 100,000 totes at a logistics site at about 98% accuracy while on task, and that its fleet has passed 65,000 operating hours. Boston Dynamics began producing the electric Atlas in January 2026 and says all of its 2026 units are committed. At Hyundai’s Georgia plant, Atlas is learning to sequence car parts for assembly, according to Electrek.
Outside observers are more cautious. The roboticist Ken Goldberg told The Verge that success rates for tasks in the lab have stalled at around 70–80%, which is not good enough for industry. IEEE Spectrum notes that many industrial customers expect 99.99% uptime, which Agility says it has shown only in specific, not general-purpose, applications.
What to watch next
- Hands: the number of active joints and how long they last in a working shift, not demo videos.
- Working runtime: battery life including reserve, and the charging time between tasks.
- Reliability: task success across thousands of repetitions, reported by someone other than the vendor.
- Cost of failure: how a robot behaves around people when something goes wrong.
Our take: capable, narrow and improving. Humanoids already do useful, repetitive work in controlled sites. The distance to a general-purpose robot is mostly about hands and reliability, and the honest way to track it is to compare working hours and success rates, not peak specs.
Questions and answers
Are humanoid robots hydraulic or electric?
The leading ones are now electric. Boston Dynamics retired its hydraulic Atlas in April 2024 and unveiled an electric one, and dropping hydraulics removes hoses and widens the range of motion. Hydraulic legged robots are also noisy and hard to scale down.
Why are robot hands so hard to build?
A hand must be dexterous, strong, durable, cheap, repairable and full of sensors at once. Fingers need to be soft for delicate objects but survive constant wear, and a machine cannot heal like skin.
How long can a humanoid robot work on one charge?
Vendors publish between about 90 minutes and 4 hours depending on the model, but the figures are measured under undisclosed workloads. One company's former product chief said that 60 of a 90-minute battery is kept as reserve.
Do robots learn to walk in the real world?
Mostly in simulation. Controllers are trained on thousands of simulated robots, then copied to a real one. ETH Zurich showed in 2019 that this works on a legged robot, with training taking under 11 hours on one desktop computer.
Do you agree?
Take it to Chevlume: write your view, add your sources and let others answer. The closed alpha is by invitation.
Keep reading
Industrial Robots by the Numbers
How many industrial robots exist, where they work, who has the most, and why density tells a different story than totals. IFR figures for 2025.
ROBOTICS / DISCOVERYShakey, the Robot That Gave Us A*
The wobbly 1960s robot that reasoned about rooms and blocks, and left behind the A* and STRIPS algorithms still used in planning.
ROBOTICS / DISCOVERYIngenuity: The Mars Helicopter
Planned for five flights, it made 72. How a helicopter with a mobile-phone processor flew in air 1% as dense as Earth's, and what ended its mission.
Sources
- An Electric New Era for Atlas · Boston Dynamics · Accessed
- Hello, Electric Atlas · IEEE Spectrum · Accessed
- Learning Agile and Dynamic Motor Skills for Legged Robots · arXiv (Hwangbo et al., ETH Zurich / Intel) · Accessed
- Atlas specifications (spec sheet PDF) · Boston Dynamics · Accessed
- Unitree H1 – Full-size universal humanoid robot · Unitree Robotics · Accessed
- Agility Unveils Digit 5 Humanoid Robot Built for Cooperatively Safe Work at Scale · Agility Robotics (press release) · Accessed
- Reality Is Ruining the Humanoid Robot Hype · IEEE Spectrum · Accessed
- Whole-Body Model-Predictive Control of Legged Robots with MuJoCo · arXiv (ICRA 2026) · Accessed
- Just because Tesla’s Optimus is in production doesn’t mean it’s actually ready · The Verge · Accessed
- Robot Hands for Modern AI and Real Work · Boston Dynamics · Accessed
- Unitree G1 humanoid robot parameters · Unitree Robotics · Accessed
- Tesla pushes Optimus deliveries to late 2027, says hands still a problem · Electrek · Accessed
- Learning agile and dynamic motor skills for legged robots · arXiv · Accessed
- Learning Quadrupedal Locomotion over Challenging Terrain · Science Robotics (accepted version, Lee, Hwangbo, Wellhausen, Koltun, Hutter) · Accessed
- Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning · arXiv (Rudin, Hoeller, Reist, Hutter) · Accessed
Published
Discuss on X