Robotics · Shuja Abidi

Quadrupeds

Robot dogs: who makes them, and how they learn to walk

Four-legged robots have left the lab. They inspect oil platforms, power stations and tunnels, and the cheapest now cost about as much as a high-end laptop. Most of today's robot dogs walk with a controller trained by reinforcement learning (RL).

1

Introduction

Four legs are a good compromise: stable enough to stand still, agile enough to run and climb.

Wheels are efficient on flat floors but stop at stairs, rubble and deep snow. Legs can step over things. With four legs a robot can always keep three on the ground if it needs to (see the gait diagram), so it is much easier to balance than a two-legged humanoid.

The modern story starts with BigDog (2005), a hydraulic, petrol-powered pack robot built by Boston Dynamics for the US military. It proved that a legged robot could walk on ice, snow and hills. Its successors became smaller, electric and quieter. Today the best-known robot dogs come from a handful of companies and labs:

A yellow Spot robot dog walking on a forest path
SpotBoston Dynamics, USAPhoto: Jonte, CC BY-SA 4.0
An ANYmal X robot climbing steep yellow stairs on an offshore platform
ANYmalANYbotics and ETH Zurich, SwitzerlandPhoto: Cvoellmy, CC BY-SA 4.0
A grey Unitree Go2 robot dog walking on snow
Go2Unitree Robotics, ChinaPhoto: HotNews Romania, CC BY 3.0
A tan Vision 60 robot walking on an airfield, with aircraft behind
Vision 60Ghost Robotics, USAPhoto: US Air Force, public domain

Two more matter a lot for research, though no free photos of them exist: MIT's Mini Cheetah, a small, fast lab robot whose software is open, and Deep Robotics' Lite3, a low-cost Chinese robot dog with open training code.

Why trained RL took over

Early robot dogs walked with hand-designed controllers: gait timers, foot-placement rules and, later, model predictive control (MPC). These work well on ground the engineers planned for. RL changed the picture around 2019–2020, when policies trained in simulation started walking real robots over mud, snow, stairs and rubble more robustly than hand-designed controllers. The reasons:

Robustness

In simulation, the policy has met millions of slips, pushes and odd surfaces before it ever touches a real floor.

Speed of development

Instead of hand-tuning a controller for each new terrain, engineers change the reward and the training worlds, then train again overnight.

Cheap to run

A trained policy is a small neural network. It runs in well under a millisecond on the robot's own computer, with no optimisation to solve.

How the training itself works is explained on the Background page.

2

Boston Dynamics

The company that made robot dogs famous, and how its approach moved from hand-built control to learning.

Boston Dynamics was founded in 1992 by Marc Raibert as a spin-off from MIT, where his Leg Lab had built hopping and running robots. For two decades its legged robots were funded mostly by US defence research (DARPA). Their videos of robots walking on ice, being kicked and getting back up made legged robots known to the public.

The large LS3 robot walking in a forest, carrying packs on its back
LS3 (Legged Squad Support System)A pack robot for soldiers, tested around 2012–2014Photo: DARPA, public domain
A yellow Spot robot walking on a dark stage lit in pink
SpotThe electric, commercial robot dog, on stage at Web SummitPhoto: Web Summit, CC BY 2.0

From BigDog to Spot

  1. Founded.Marc Raibert starts Boston Dynamics, bringing the MIT Leg Lab's ideas about dynamic balance into a company.
  2. BigDog.A hydraulic pack robot powered by a petrol engine, funded by DARPA. It walks on rough ground, snow and ice, and recovers when kicked. It is loud, but it proves a legged robot can work outdoors.
  3. LS3.A bigger version meant to carry soldiers' gear, tested with the US Marines. The project was later set aside, partly because the engine was too loud for patrols.
  4. Smaller and electric.New, smaller dogs (Spot, then the 25 kg SpotMini) drop the petrol engine. The quieter electric design becomes the basis of the product.
  5. New owners.Google buys the company in 2013, SoftBank in 2017, and Hyundai Motor Group takes control in 2021.
  6. Spot goes on sale.First to early customers in 2019, then to anyone in June 2020. It's used for inspecting plants and sites, for public safety and for research.
  7. Learning joins the controller.Boston Dynamics adds reinforcement learning to Spot's walking. The RAI Institute, a research lab Raibert started in 2022, shows an RL policy running Spot at over 5.2 m/s.

How the control changed

Model-based, for most of its history. Boston Dynamics' robots balanced with controllers built by engineers from physics: a model of the body, rules for where to place the feet, and later optimisers that plan foot forces a short time ahead (MPC). This gave the smooth, reliable walking in their famous videos, but every new behaviour took a lot of expert tuning.

Model-based plus RL, today. Rather than throw the old controller away, Boston Dynamics added a learned policy to Spot's locomotion system. The model-based parts that already work well stay, and the policy handles what is hard to write rules for, such as slips and unusual surfaces. It is a good example of the hybrid design described on the Background page.

Open or proprietary?

Proprietary The robots, their walking controllers and firmware are closed.

  • Spot SDK: the code for controlling Spot from your own programs is public on GitHub, but under Boston Dynamics' own SDK licence, not an open-source licence.
  • RL research: researchers can train their own Spot policies in NVIDIA Isaac Lab and run them on the robot through Boston Dynamics.

Boston Dynamics also builds Atlas, a humanoid, which belongs on the Humanoid page.

3

At a glance

Who uses RL, and what you can download.

RobotMakerWalking controllerRL code you can useOverall
SpotBoston DynamicsModel-based, with RL added to itA research kit to train and run your own policiesProprietary
ANYmalANYbotics / ETH ZurichRL (in ETH's research)ETH's training code, legged_gym (BSD-3)Mixed
Go2Unitree RoboticsBuilt-in controller (closed)Unitree's own RL training code (BSD-3 and others)Mixed
Mini CheetahMITMPC; RL in later researchFull controller software (MIT licence)Open
Lite3Deep RoboticsRLOfficial training and deployment code (BSD-3)Mixed
Vision 60Ghost RoboticsNot publicly documentedNone publicProprietary

"Mixed" means the robot and its built-in software are closed, but the maker or its partner lab 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.

Spot robot dog on a forest path

Spot

Boston Dynamics, USA · on sale since 2020 · industrial inspection, public safety, research

Controller
Spot launched with a model-based controller: a physics model of the robot, and an optimiser that plans how the feet push. Its default top speed is about 1.6 m/s.
How RL is used
Boston Dynamics has since added RL to Spot's locomotion system. The model-based parts that already work well are kept, and a learned policy takes over the harder decisions, such as reacting to slippery ground. Separately, the RAI Institute trained an end-to-end RL policy that ran a real Spot at over 5.2 m/s, about three times its default speed.
Why RL
Better handling of slips, trips and unusual surfaces that are hard to write rules for, plus higher speed.
Open or closed
ProprietaryThe robot's software is closed. Boston Dynamics offers researchers a way to train their own policies in NVIDIA Isaac Lab and run them on Spot, but the built-in controller isn't released.
ANYmal X on offshore platform stairs

ANYmal

ANYbotics (an ETH Zurich spin-off), Switzerland · industrial inspection, including offshore and explosive-risk areas

Controller
ANYmal is the robot on which much of today's learned locomotion was first shown, by ETH Zurich's Robotic Systems Lab.
How RL is used
2019: a policy trained in simulation walked the real robot, helped by a neural network that copies how its motors behave. 2020: an RL policy walked over mud, snow, rubble and running water outdoors. 2022: a policy that combines what the legs feel with what the cameras see hiked a mountain trail.
Why RL
Industrial sites are full of stairs, gratings and clutter. RL policies stayed upright across far more of these surfaces than the hand-designed controllers before them.
Open or closed
MixedThe commercial robot's software is proprietary. ETH's research training code, legged_gym, is open (BSD-3 licence) and became a starting point for many other labs and companies.
Unitree Go2 walking on snow

Go2

Unitree Robotics, China · launched 2023 · education, research, hobby, and mountain rescue trials

Controller
The Go2 comes with Unitree's built-in walking controller, which users drive with speed commands through Unitree's SDK.
How RL is used
Unitree publishes its own RL training code, so anyone can train a new walking policy for the Go2 in simulation and run it on the robot in place of the built-in one. Many university labs do exactly this.
Why RL
A low price plus open training code makes the Go2 one of the most common robots for RL locomotion research.
Open or closed
MixedThe built-in controller and firmware are closed. The training code is open: unitree_rl_gym (BSD-3, Isaac Gym), plus newer versions for Isaac Lab and MuJoCo on Unitree's GitHub.
No freely licensed photo of the Mini Cheetah is available.

Mini Cheetah

MIT Biomimetic Robotics Lab, USA · 2019 · research

Controller
A small, light robot built for agility; it was shown doing backflips. Its original controller uses convex MPC to plan foot forces, the same approach as the larger MIT Cheetah 3.
How RL is used
In 2022, MIT's Improbable AI Lab trained an RL policy in simulation that ran the Mini Cheetah at up to 3.9 m/s on grass, ice and gravel (Margolis et al., RSS).
Why RL
To push speed past what the model-based controller reached, on real outdoor ground.
Open or closed
OpenThe full controller software, Cheetah-Software, is released under the MIT licence. That makes it one of the few robot dogs whose complete model-based controller can be studied line by line.
No freely licensed photo of the Lite3 is available.

Lite3

Deep Robotics, China · education and research

Controller
A low-cost robot dog in the same class as the Go2.
How RL is used
Deep Robotics publishes an official RL pipeline: training in NVIDIA Isaac Lab, then sim-to-sim checks and deployment on the real robot with an ONNX policy.
Why RL
Like Unitree, the company makes it easy for customers to train their own walking policies, which is what research labs want.
Open or closed
MixedTraining code rl_training is open (BSD-3), with a separate deployment repository on Deep Robotics' GitHub. The robot's built-in software is closed.
Vision 60 robot on an airfield

Vision 60

Ghost Robotics (a University of Pennsylvania spin-off), USA · defence and security, rugged outdoor use

Controller
Built for hard outdoor work, with a metal body that survives falls. It uses joint sensors and an IMU to feel the ground, plus cameras.
How RL is used
Ghost Robotics has not published how its walking controller works, so whether it uses trained RL is not public.
Why it's here
It's a widely deployed robot dog, and a reminder that not every company shares its methods.
Open or closed
ProprietaryNo public controller or training code.

5

Open source or proprietary?

"Open" usually covers the training code, not the robot.

A robot dog has several layers of software. For almost every commercial robot, the lowest layers, the motor firmware and the built-in walking controller, are proprietary: you can use them but not see or change them. What companies increasingly open up is the training pipeline: robot models, simulation environments, reward settings and scripts to deploy a trained policy.

This split suits both sides. Companies keep their tuned product controller private, but researchers can still train their own policies and run them on the real robot through a low-level interface that sends joint targets directly.

What the licences mean

  • MIT and BSD-3: use, change and even sell the code, as long as you keep the copyright notice. BSD-3 also forbids using the authors' names to promote your product.
  • Apache-2.0: similar, and adds an explicit patent grant.
  • Proprietary: no rights beyond what the maker's terms allow.

What you can usually get

  • Often open Robot model files for simulators
  • Often open RL training environments and rewards
  • Often open Deployment scripts for a trained policy
  • Sometimes Pretrained policies
  • Rarely open The built-in product controller
  • Rarely open Motor firmware