Robotics · Shuja Abidi

Humanoid

Humanoids: who makes them, and how they learn to walk

For decades, humanoids walked carefully, with knees bent, following motions planned by engineers. In just a few years, reinforcement learning (RL) has given them a natural stride, push recovery and even dance. Today almost every new humanoid learns to walk in simulation.

1

Introduction

Why build a robot shaped like a person? Because our world is shaped for people.

Stairs, doors, tools, shelves and vehicles are all designed for human bodies. A humanoid could, in principle, work anywhere a person can without changing the building. The price is balance: two legs, a high body and small feet make a humanoid far harder to keep upright than a robot dog (see humanoid motion).

Honda's ASIMO (2000) showed that a humanoid could walk smoothly, using motions planned to keep the zero moment point inside its feet. Boston Dynamics' Atlas (2013) brought dynamic balance, and later parkour and backflips. Since about 2023, a wave of companies, many in the US and China, have started building humanoids for factories, warehouses and research.

Honda's white ASIMO humanoid standing on display
ASIMOHonda, Japan · 2000Photo: Syced, CC0
The electric Atlas humanoid standing on a display stand
Atlas (electric)Boston Dynamics, USA · 2024Photo: Damian B Oh, CC BY-SA 4.0
A Unitree G1 humanoid standing in a crowded exhibition hall
G1Unitree Robotics, China · 2024Photo: Sayanesy, CC0
A white and black Tesla Optimus humanoid waving among people at an evening event
OptimusTesla, USAPhoto: Steve Jurvetson, CC BY 2.0

Others matter just as much but have no free photos: Figure's humanoids, Agility Robotics' Digit (already working in warehouses), Unitree's full-size H1, and small open research robots such as the Booster T1 and Berkeley Humanoid Lite.

Why trained RL matters even more for humanoids

Balance is harder

Hand-designed humanoid controllers kept the robot careful and slow so its simple models stayed valid. A trained policy has learned to recover from thousands of slips and pushes in simulation, so it can walk faster and more naturally.

Many more joints

A humanoid has 20 to over 40 motors, including arms and waist. Writing rules for all of them together is very hard; a neural network learns to coordinate them.

Learning from people

Rewards can ask the robot to copy recordings of human motion. That's how new humanoids get heel strikes, toe-offs and arm swing that look human.

2

Boston Dynamics Atlas

The humanoid that set the bar for dynamic balance, rebuilt as an electric robot that learns.

The original hydraulic Atlas humanoid, with exposed metal frame and hoses
Atlas, hydraulic (2013)Built for the DARPA Robotics ChallengePhoto: DARPA, public domain
The electric Atlas humanoid on a display stand
Atlas, electric (2024)Lighter, electric motors, joints that rotate further than a person'sPhoto: Damian B Oh, CC BY-SA 4.0

From research robot to product

  1. Hydraulic Atlas.Built for DARPA's Robotics Challenge, where university teams programmed it for disaster-response tasks such as opening doors and turning valves.
  2. Dynamic and athletic.Newer, more compact versions walk in snow, recover from shoves, and later run parkour courses and do backflips, all with model-based control.
  3. Electric Atlas.The hydraulic Atlas retires. Its electric successor is aimed at real work, starting in Hyundai's car factories.
  4. Large behavior models.A partnership with Toyota Research Institute brings in AI models, similar in spirit to large language models, that learn whole tasks from demonstrations.
  5. Reinforcement learning.A partnership with the RAI Institute focuses on RL and moving skills from simulation to the real robot, using the same approach that made Spot run at 5.2 m/s.

How the control changed

Model-based, for the hydraulic Atlas. Its famous parkour and backflips came from model predictive control: engineers designed each move, and an optimiser worked out, many times a second, how to use the whole body to land it. It was impressive, but every new skill took months of expert work.

Learning, for the electric Atlas. The new robot is built to learn: RL in simulation for robust movement, and large behavior models for whole tasks taught by demonstration. The goal is new skills in days, not months.

Open or proprietary?

Proprietary Atlas, its controllers and its learned models are closed. Unlike Spot, Atlas is not sold to researchers, so there is no public SDK.

3

At a glance

Who uses RL, and what you can download.

RobotMakerWalking controllerRL code you can useOverall
ASIMOHondaModel-based (ZMP); retiredNoneProprietary
AtlasBoston DynamicsModel-based, now RL and learned modelsNoneProprietary
G1, H1Unitree RoboticsBuilt-in controller (closed)Training code and trained policies (BSD-3)Mixed
Figure 02Figure AIEnd-to-end RLNoneProprietary
DigitAgility RoboticsRL whole-body controllerNoneProprietary
OptimusTeslaRL trained in simulationNoneProprietary
T1Booster RoboticsRLBooster Gym (Apache-2.0)Mixed
Berkeley Humanoid LiteUC BerkeleyRLHardware, code and training, all releasedOpen

"Mixed" means the robot and its built-in software are closed, but the maker has released training code. Licences can change, so check each repository before reuse.

4

Each robot and its RL

What the robot is for, how trained RL is used on it, why, and how open it is.

Unitree G1 humanoid in an exhibition hall

Unitree G1 and H1

Unitree Robotics, China · H1 (full size) 2023, G1 (about 1.3 m) 2024 · research, education, shows

Controller
Both ship with Unitree's built-in walking controller, driven through Unitree's SDK.
How RL is used
Unitree publishes RL training code for G1, H1 and H1-2, together with trained walking policies. A policy reads the robot's tilt, joint angles and speeds and the speed command, and outputs a target angle for each leg joint about 50 times a second.
Why RL
It makes a low-cost humanoid that labs can retrain, which is why the G1 has become one of the most common research humanoids.
Dance and kung fu
Unitree's famous dancing and kung fu videos are also RL, trained to copy human motion capture. A few example dance policies are released; the show demos are not. See Dancing and fighting demos.
Open or closed
MixedThe built-in controller and firmware are closed. unitree_rl_gym is open (BSD-3), with newer Isaac Lab and MuJoCo versions on Unitree's GitHub.
No freely licensed photo of Figure's robots is available.

Figure 02

Figure AI, USA · general-purpose humanoid, tested in car manufacturing

Controller
One end-to-end neural network walks the robot.
How RL is used
The policy is trained entirely in a GPU-accelerated simulation, with years' worth of practice in a few hours. Its reward asks it to follow recordings of human walking, which gives heel strikes, toe-offs and arm swing. Domain randomisation and fast torque control on the robot help it transfer. Figure showed ten robots walking with the same policy, unchanged.
Why RL
A natural, efficient walk that works across many robots and floors without tuning each one.
Open or closed
ProprietaryDescribed on Figure's blog; no code released.
No freely licensed photo of Digit is available.

Digit

Agility Robotics, USA · moving totes and boxes in warehouses

Controller
Digit grew out of Cassie, Agility's earlier pair of legs. Cassie was a favourite platform for RL research in universities, and with an RL policy it set a Guinness record for the fastest 100 m by a bipedal robot in 2022.
How RL is used
A 2021 research paper showed the first learned policy moved from simulation to a real Digit. Agility now describes a whole-body control "foundation model" for Digit, trained with RL in NVIDIA's simulator and checked in a second simulator (MuJoCo) so it doesn't depend on one physics engine.
Why RL
A warehouse robot must stay safe and stable while carrying loads of different weights, all day.
Open or closed
ProprietaryDescribed on Agility's site; no code released.
Tesla Optimus humanoid at an event

Optimus

Tesla, USA · intended for factory work, then homes

Controller
Tesla has said little in detail about how Optimus is controlled.
How RL is used
Tesla's Optimus team said a dance routine was trained entirely in simulation with RL and then moved to the real robot. Its job adverts describe end-to-end RL policies for walking, balance and recovery. Details of the walking controller itself are not public.
Why RL
Fast, whole-body skills that would take a long time to program by hand.
Open or closed
ProprietaryNo public controller, models or code.
No freely licensed photo of the T1 is available.

T1

Booster Robotics, China · small humanoid for research and robot football

Controller
A learned walking policy, trained with the company's own open framework.
How RL is used
Booster Gym covers the whole route: training in Isaac Gym with PPO and domain randomisation, a sim-to-sim check in MuJoCo, then deployment on the real robot with no extra tuning. It walks in any direction, resists pushes and adapts to terrain.
Why RL
Like Unitree, Booster makes it easy for labs to train their own policies.
Open or closed
Mixedbooster_gym is open (Apache-2.0); the robot's built-in software is closed.
No freely licensed photo of Berkeley Humanoid Lite is available.

Berkeley Humanoid Lite

UC Berkeley, USA · 2025 · a fully open humanoid for research and teaching

Controller
A small (about 0.8 m), mostly 3D-printed humanoid that can be built for under US$5,000.
How RL is used
Its walking policy is trained with RL in simulation and runs on the real robot with no further tuning.
Why RL
To show that learned walking works even on cheap, 3D-printed hardware that anyone can build.
Open or closed
OpenThe hardware design, robot code, and training and deployment tools are all released, on the project site. That makes it one of the very few humanoids that are open from motors to policy.

5

Dancing and fighting demos

Unitree's videos of dancing, kung fu and boxing robots: how they're made, and what you can download.

Walking only needs a speed command. A dance or a kung fu routine is different: the robot must copy a whole sequence of human body poses while staying balanced. The method behind most of these demos is called motion imitation (or motion tracking).

  1. Record a person. A dancer or martial artist performs in a motion-capture suit, or is filmed on video.
  2. Fit it to the robot. The human motion is "retargeted" to the robot's body, which has different limb lengths and fewer joints.
  3. Train in simulation. An RL policy is rewarded for matching the recorded pose at every moment without falling. Domain randomisation helps it survive the real world.
  4. Run it on the robot. The policy reads the robot's state and the next target pose, and outputs joint targets about 50 times a second.

What Unitree has released

Checked in Unitree's GitHub repositories:

WhereWhat you getLicence
unitree_rl_lab
Isaac Lab
Two trained dance policies for the G1, ready to deploy on the robot: Dance 102 and Gangnam Style. Each comes with the motion it copies, and the deployment code has a "Mimic" mode to play them.Apache-2.0
unitree_rl_mjlab
MuJoCo
A motion-imitation training task for the G1 (29 and 23 joint versions), one trained dance policy, and the motion file it learned from. Bring your own motion file to train a new routine.Apache-2.0
Motion dataHuman motion-capture clips from the LAFAN1 dataset, retargeted by Unitree to fit its humanoids. The original LAFAN1 recordings come from Ubisoft, under CC BY-NC-ND 4.0: free for research, not for commercial use.Non-commercial

What Unitree has not released

So: the method is open and you can train your own dance with it, but Unitree's best show routines stay private.

Open alternatives others have released for the G1

Several research groups have released complete motion-imitation systems that run on a real G1. Unitree's own imitation code points to the first one.

BeyondMimic

UC Berkeley, 2025

Tracks highly dynamic human motions, such as sprints, jumps and crawling. The paper reports 30 motions transferred to a real G1 without per-motion tuning, and adds a diffusion model to combine skills.

MIT whole_body_tracking

ASAP

Carnegie Mellon University and NVIDIA · RSS 2025

Teaches agile, athlete-style moves. It first trains in simulation, then learns a correction from real-robot data to close the gap between simulated and real physics.

MIT LeCAR-Lab/ASAP

ProtoMotions

NVIDIA

A GPU-accelerated framework for training simulated humans and humanoid robots. It includes a documented G1 workflow, from motion data to a real robot.

Apache-2.0 NVlabs/ProtoMotions

6

Open source or proprietary?

Humanoids are more closed than robot dogs, but the gap is narrowing.

The biggest humanoid companies, including Boston Dynamics, Figure, Agility and Tesla, keep their controllers and models private. They are racing to sell robots for factories and warehouses, and their learned walking and manipulation are what they compete on.

The open side comes from two directions. Low-cost makers such as Unitree and Booster publish their RL training code so universities will buy and build on their robots. University labs such as Berkeley release entire robots, from the 3D-printed parts to the trained policy.

For anyone learning, that's good news: you can train a humanoid walking policy today with the same open tools the companies started from, and test it in a free simulator such as MuJoCo before touching real hardware.

What you can usually get

  • Often open Robot model files for simulators
  • Often open RL training code from low-cost makers
  • Sometimes Pretrained walking policies
  • Rarely Full hardware designs (research robots)
  • Closed Controllers of the big commercial humanoids

The licence meanings (MIT, BSD-3, Apache-2.0) are explained on the Quadrupeds page.