CtrlK
A quadruped robot is a four-legged robotic system that uses dynamic, articulated legs to walk, trot, run, climb, and traverse uneven terrain. By emulating the locomotion strategies of animals such as dogs, goats, and horses, quadrupeds can reach environments that wheeled or tracked machines cannot, making them valuable for industrial inspection, military reconnaissance, search and rescue, construction monitoring, and research. Modern quadrupeds combine actuated mechanical legs, inertial and visual sensing, real-time control electronics, and increasingly learned reinforcement learning policies trained in simulation. Once exotic and prohibitively expensive, quadruped platforms became broadly accessible in the early 2020s through commercial offerings from Boston Dynamics, ANYbotics, Unitree, Deep Robotics, Ghost Robotics, and Xiaomi, with consumer-grade models now selling for under $2,000.[1][2]
Quadruped robots sit at the intersection of mechanical engineering, control theory, and modern machine learning. They are widely viewed as a practical proving ground for robotics research before techniques migrate to bipedal platforms such as the humanoid robot, and many of the algorithms now powering Tesla Optimus and other commercial humanoids were first validated on four-legged systems.
A quadruped robot typically has four limbs, each containing two or three powered joints (a hip in the sagittal plane, a hip in the frontal plane, and a knee), giving each leg between 8 and 12 degrees of freedom across the whole body. The legs use either electric motors with planetary or harmonic gearboxes (the dominant approach since around 2015) or hydraulic actuators (used on early Boston Dynamics platforms such as BigDog and the Atlas robot family). Onboard electronics include an inertial measurement unit, joint torque or position encoders, a forward-facing depth camera or LiDAR, and one or more compute modules running real-time control loops at 500 Hz to 1 kHz.
Unlike static-stability legged platforms (such as hexapods or six-legged walkers), modern quadrupeds are dynamically stable, meaning they continually fall and catch themselves the way galloping animals do. This dynamic balance was the key conceptual contribution of Marc Raibert's MIT Leg Lab in the 1980s and remains foundational in the field. Dynamic stability allows quadrupeds to traverse stairs, debris, mud, snow, and slopes that would topple a statically balanced machine.[3]
The transition from classical model-predictive control (MPC) to learning-based controllers around 2018 to 2020 dramatically expanded the terrain quadrupeds can handle. Today, the most capable controllers are neural networks trained in simulation through deep reinforcement learning and then deployed on hardware via sim-to-real transfer.
The first widely cited four-legged powered walker was the General Electric Walking Truck, designed by Ralph Mosher and unveiled in 1968 under a U.S. Army Tank-Automotive Research Center contract. The 3,000-pound, 11-foot hydraulic quadruped was operated by a human driver via hand and foot pedals with rudimentary force feedback. It could carry 500 pounds, climb two-foot obstacles, and walk at five miles per hour, but was never autonomous.[4]
In the early 1980s at Carnegie Mellon University, Ivan Sutherland built the Trojan Cockroach (sometimes called the Sutherland Trotting Machine), a DARPA-funded six-legged hydraulic walker. Although a hexapod rather than a quadruped, it was the first computer-controlled walking machine that could carry a human rider and is recognized as a foundational milestone for legged robotics.
The true intellectual breakthrough came from Marc Raibert, who founded the CMU Leg Laboratory in 1980 and moved it to MIT in 1986. Raibert showed that legged locomotion could be decomposed into three decoupled control laws: forward speed, body attitude, and hopping height. He started with one-legged hoppers, progressed to two-legged hoppers, and built four-legged trotting and bounding machines that demonstrated true dynamic balance. In 1992, Raibert spun out Boston Dynamics to commercialize the lab's research.[3][5]
In 2005, Boston Dynamics, working with Foster-Miller, the NASA Jet Propulsion Laboratory, and Harvard's Concord Field Station, unveiled BigDog, the first widely publicized rough-terrain quadruped. Funded by DARPA's Maximum Mobility and Manipulation program, BigDog was three feet long, weighed 240 pounds, and was powered by a 15-horsepower two-stroke gasoline engine driving four hydraulic cylinder-actuated legs. It could carry 340 pounds, traverse 35-degree slopes, run at 4 miles per hour, and recover from being kicked sideways on ice in viral demonstration videos. BigDog set the public mental image of what a robot dog looked like.[1][6]
Boston Dynamics followed BigDog with a research-focused LittleDog platform around 2010, then AlphaDog (also known as the Legged Squad Support System or LS3) in 2012. Funded by DARPA and the U.S. Marine Corps with a $32 million contract, LS3 was a refined and quieter cargo-carrying quadruped designed to follow Marine squads through rough terrain carrying up to 400 pounds for 20 miles without refueling. The Marines tested LS3 in field exercises in Hawaii in 2014, but ultimately rejected the platform because the gas engine was too loud for tactical operations. The program was shelved in 2015.
A major shift began in 2013 to 2015 when MIT's Biomimetic Robotics Lab, led by Sangbae Kim, demonstrated that high-torque electric motors could match or exceed the efficiency of hydraulic actuators for legged locomotion. The MIT Cheetah (2009) and Cheetah 2 (2013) showed energy-efficient running, and Cheetah 3 (2018) was a robust 90-pound platform capable of climbing stairs without sight. In 2019, MIT released the Mini Cheetah, a 20-pound, $10,000 educational quadruped that became the first four-legged robot to perform a 360-degree backflip from a standing start. Mini Cheetah was distributed to research labs around the world and seeded an entire generation of quadruped researchers.[7]
In parallel, ETH Zurich's Robotic Systems Lab (RSL) and Autonomous Systems Lab (ASL) developed ANYmal, a series-elastic actuator-based quadruped designed for autonomous industrial inspection. ANYmal won the ARGOS Challenge organized by Total in 2017, demonstrating navigation of a multi-floor offshore oil and gas platform replica. The technology was spun out as ANYbotics in 2016, which now sells the ANYmal C and ANYmal D platforms commercially and operates the explosion-proof ANYmal X for hazardous chemical and oil and gas environments (the world's first ATEX-certified legged robot).[2]
In June 2015, Boston Dynamics introduced Spot, an electric quadruped weighing about 75 pounds that replaced BigDog's hydraulics and gasoline engine with quiet electric actuators and lithium batteries. Spot was iterated through several generations (Spot Classic, SpotMini, and the production Spot in 2019). On June 16, 2020, Boston Dynamics began commercial sales of Spot at a starting price of $74,500 (excluding accessories such as the $29,750 Spot Arm and $4,620 spare batteries). Spot became the first commercially available autonomous quadruped sold to the general public.[8]
The second great inflection point came from machine learning. In January 2019, Jemin Hwangbo and colleagues at ETH Zurich published "Learning agile and dynamic motor skills for legged robots" in Science Robotics. The paper trained a neural network policy in the RaiSim physics simulator and transferred it directly to ANYmal hardware. The learned controller followed velocity commands more efficiently than hand-engineered baselines and could autonomously recover from a fall, including legs-up. The work established the now-standard sim-to-real recipe: massively parallel simulation, domain randomization, and an actuator network learned from real motor data.[9]
In October 2020, Joonho Lee and colleagues followed up with "Learning quadrupedal locomotion over challenging terrain." The blind teacher-student controller used a temporal convolutional network over proprioceptive history to take ANYmal through mud, snow, dense vegetation, gushing water, and broken rubble that no prior published controller could handle, demonstrating zero-shot generalization to terrains never seen in training.[10]
In January 2022, Takahiro Miki and colleagues extended this with "Learning robust perceptive locomotion for quadrupedal robots in the wild," fusing proprioception with depth-camera input via an attention-based recurrent encoder. Four ANYmal robots traversed over 1,700 meters of underground tunnels, urban sites, and natural caves without a single fall, and one completed an Alpine hike at human pace.[11] These three papers established RL as the dominant paradigm for quadrupeds and influenced subsequent agility_robotics bipeds, humanoids, and dexterous manipulators.
From 2021 onward, Chinese manufacturers drove quadruped prices down an order of magnitude. Unitree Robotics, founded in Hangzhou in 2016 by Wang Xingxing, released the Laikago (2017), AlienGo and A1 (2019), Go1 (2021, from $2,700), B1 (2021), the industrial B2 (2023), and the Go2 in mid-2023 (from $1,600 Air, $2,800 Pro). The Go2 ships with a 4D LiDAR with a 360-by-96-degree field of view, learned gaits including upside-down walking, and 1 to 2 hours of battery life. Deep Robotics, founded in 2017 from Zhejiang University, sells the inspection-focused X20 (IP66, 4-hour battery), the larger X30, and the Lynx M20, the world's first wheeled-legged quadruped capable of 5 m/s while retaining stair-climbing. Xiaomi released CyberDog in 2021 and CyberDog 2 in August 2023. The 8.9 kg CyberDog 2 runs an NVIDIA Jetson Xavier NX, integrates 19 sensors, performs continuous backflips, and retails for around $1,400.[12][13]
Quadrupeds have become a primary deployment target for vision-language-action policies. NVIDIA's Project GR00T, announced in March 2024 for humanoids and expanded as Isaac GR00T N1 in March 2025, has been adapted for quadruped manipulation. Boston Dynamics has demonstrated learned-behavior models on Spot for pick-and-place tasks, and Unitree, Deep Robotics, and ANYbotics have added LLM interfaces for natural-language task specification. As of 2026, the line between research quadrupeds and general-purpose embodied AI platforms has largely dissolved.
| Robot | Manufacturer | Year | Weight | Top speed | Notable feature | Indicative price |
|---|---|---|---|---|---|---|
| Walking Truck | General Electric | 1968 | 1,400 kg | 8 km/h | First powered four-legged walker | Research only |
| MIT Quadruped | MIT Leg Lab | 1984 | 38 kg | 2.2 m/s | First dynamic-balance trotting robot | Research only |
| BigDog | Boston Dynamics | 2005 | 110 kg | 6.4 km/h | Hydraulic, gas-engine, kick-recovery | DARPA-funded |
| LittleDog | Boston Dynamics | 2010 | 3 kg | 0.5 m/s | DARPA learning testbed | Research only |
| AlphaDog/LS3 | Boston Dynamics | 2012 | 590 kg | 11 km/h | Marine cargo carrier, 400 lb load | $32M program |
| MIT Cheetah 3 | MIT Biomimetic Lab | 2018 | 41 kg | 3 m/s | Stair climbing, blind locomotion | Research only |
| ANYmal C | ANYbotics | 2019 | 50 kg | 1 m/s | Industrial inspection, IP67 | ~$150,000 |
| Mini Cheetah | MIT | 2019 | 9 kg | 2.45 m/s | First quadruped backflip | $10,000 (research) |
| Stanford Doggo | Stanford Robotics | 2019 | 5 kg | 0.9 m/s | Open-source, sub-$3,000 BOM | DIY |
| Spot | Boston Dynamics | 2020 | 32.7 kg | 1.6 m/s | First commercial quadruped | $74,500 |
| Vision 60 | Ghost Robotics | 2020 | 51 kg | 2.2 m/s | Military/defense focused | Classified pricing |
| Unitree Go1 | Unitree | 2021 | 12 kg | 4.7 m/s | First sub-$3,000 consumer quadruped | $2,700 |
| ANYmal X | ANYbotics | 2022 | 70 kg | 1 m/s | First Ex-certified legged robot | ~$200,000+ |
| CyberDog | Xiaomi | 2021 | 12 kg | 3.2 m/s | Open-source, NVIDIA Jetson Xavier | $1,540 |
| Unitree B2 | Unitree | 2023 | 60 kg | 6 m/s | 120 kg payload, industrial | ~$100,000 |
| Unitree Go2 | Unitree | 2023 | 15 kg | 3.5 m/s | 4D LiDAR, learned gaits | $1,600 |
| CyberDog 2 | Xiaomi | 2023 | 8.9 kg | 1.6 m/s | 19 sensors, biomimetic | $1,400 |
| ANYmal D | ANYbotics | 2023 | 50 kg | 1.2 m/s | Production inspection model | ~$150,000 |
| Deep Robotics X30 | Deep Robotics | 2023 | 56 kg | 4 m/s | IP67, 4 hour endurance | ~$60,000 |
| Lynx M20 | Deep Robotics | 2024 | 38 kg | 5 m/s | Wheeled-legged hybrid | ~$30,000 |
Hydraulic actuators (used on BigDog, LS3, and the original Atlas) deliver power density and shock tolerance but require pumps, oil reservoirs, and noisy combustion engines, and have largely fallen out of favor.
Series-elastic electric actuators, pioneered on ANYmal, place a torsional spring between motor output and joint to absorb impacts and estimate torque, giving ANYmal its compliant gait.
Quasi-direct-drive (QDD) actuators, developed at MIT and now used on Mini Cheetah, Stanford Doggo, and Unitree platforms, pair a high-torque brushless motor with a low (6:1 to 9:1) planetary gearbox. The low gear ratio preserves backdrivability so the leg senses contact through motor current alone. QDD is the de facto standard for sub-50 kg quadrupeds.
Sensing combines joint encoders, current sensors, foot-contact sensors, and a body IMU at 200 to 1000 Hz. Exteroception ranges from stereo and RGB-D cameras on research platforms to 360-degree LiDAR (Ouster, Velodyne, Livox), thermal cameras, and gas detectors on inspection-grade industrial robots.[2]
Low-level joint control runs on a real-time microcontroller at 500 Hz to 1 kHz. Planning, perception, and learned policies run on an NVIDIA Jetson Orin Nano, Orin NX, or AGX Orin module, or for ANYmal an Intel NUC. CUDA support makes the Jetson family the default choice for deep learning inference on Unitree, Deep Robotics, and Xiaomi platforms.
Classical quadruped control divides the problem into gait scheduling, swing-leg trajectory generation, and stance-leg force optimization. Boston Dynamics' published BigDog and Spot controllers used a virtual model of the body coupled to whole-body inverse dynamics. MIT's Cheetah series introduced convex Model Predictive Control (MPC) over the centroidal dynamics of the robot, solved at 30 to 100 Hz, that enables fast, adaptive trotting. ANYmal's classical controller used Whole-Body Control combined with a Riccati-based linear quadratic regulator. These methods produce smooth gaits but require careful hand-tuning and assume relatively flat ground.
Deep RL controllers train a neural network in simulation through proximal policy optimization (PPO) or similar algorithms, with the policy mapping observations (joint positions, velocities, base orientation, and command) to joint position or torque targets. Sim-to-real transfer is achieved through:
| Technique | Purpose |
|---|---|
| Massively parallel simulation | Train on thousands of robots in parallel (Isaac Lab, RaiSim) |
| Domain randomization | Vary mass, friction, motor parameters to expose policy to uncertainty |
| Actuator network | Learn the real motor's torque response from hardware data |
| Privileged learning | Train a teacher with full state, distill to a deployable student |
| Curriculum learning | Gradually increase terrain difficulty during training |
| Symmetric augmentation | Mirror left and right legs to halve data requirements |
Training now takes minutes to hours on a single GPU, and hardware deployment is essentially free of further tuning. The dominant simulators for quadruped RL are RaiSim, NVIDIA Isaac Gym/Lab, MuJoCo MJX, and Genesis.
The newest direction layers a vision-language-action model on top of a learned low-level controller. The high-level model converts a natural language instruction ("go inspect the gauge near the third valve") into a sequence of velocity commands or waypoints, which the low-level RL policy executes. NVIDIA Isaac GR00T N1, originally aimed at humanoids, has been adapted for quadrupeds by several research groups, and Boston Dynamics' Spot now ships with optional integrations for OpenAI's API for natural-language task specification.
The largest commercial market is autonomous inspection of process plants. ANYbotics has deployed ANYmal on offshore platforms, refineries, chemical plants, power stations, and underground mines for operators such as Petronas, BASF, OMV, and Shell. Spot has been used by BP on the Mad Dog deepwater platform in the Gulf of Mexico starting in 2020 and on Aker BP's Skarv installation in the Norwegian Sea, where it autonomously inspects gauges, listens for leaks, and reads meters previously checked by human technicians. National Grid, Enel, and Equinor have rolled out Spot fleets across substations and gas-processing facilities.
Spot, ANYmal, and Deep Robotics platforms capture daily 3D scans on construction sites, compare progress against BIM models, and inspect formwork and rebar placement. Pomerleau, Skanska, and Foster + Partners have publicly piloted Spot for this purpose, walking programmed routes at the end of each work day to build as-built point clouds.
Ghost Robotics Vision 60 (Q-UGV, Quadrupedal Unmanned Ground Vehicle) is the dominant Western military quadruped. The U.S. Air Force began deploying Vision 60 at Tyndall Air Force Base in 2020, and it has since been adopted by the U.S. Department of Homeland Security, the Australian Defence Force, the British Army, and over a dozen other militaries. The 51 kg V60 carries up to 10 kg of payload, walks at up to 2.2 m/s, and is rated for dust and rain. In 2023 and 2024, Ghost Robotics demonstrated weaponized variants armed with rifles or 6.5 mm sniper systems, igniting public debate about lethal autonomous robots. The U.S. Army has signed multiple V60 contracts across logistics, perimeter security, and explosive ordnance disposal.
The NYPD, Massachusetts State Police, LAPD, and Honolulu Police Department have deployed Spot for SWAT and bomb-squad reconnaissance, including active-shooter standoffs and hazardous-materials assessment. Use is controversial: the NYPD deployed Spot in 2020, retired it in 2021 after public backlash, and reinstated a fleet under more transparent rules in 2023.
Quadrupeds excel in collapsed buildings, mine shafts, and tunnel systems. ANYmal and Spot placed prominently in the DARPA Subterranean Challenge (2018 to 2021). After the February 2023 Turkey-Syria earthquake, Boston Dynamics deployed Spot units to inspect partially collapsed structures, and Deep Robotics has supplied X30 units to Chinese fire-rescue services. Industry estimates put more than 12,000 quadruped units deployed for search-and-rescue tasks globally in 2023.[14]
Boston Dynamics' viral marketing for Spot, including the 2020 "Do You Love Me?" dance video and a Christmas dance with Atlas, destigmatized quadrupeds in popular culture. Spot has appeared in K-pop music videos, Cirque du Soleil performances, Hyundai commercials, and at Disney parks. Unitree Go robots have appeared at the China Central Television Spring Festival Gala, including a synchronized 16-robot performance in 2025.
Low-cost platforms such as Mini Cheetah, Stanford Doggo (open-source under $3,000), Unitree Go1 EDU, and the MIT Pupper are now standard hardware in graduate robotics courses. Stanford Doggo, released in 2019 by Nathan Kau's Extreme Mobility student group with full BOM, CAD, and firmware on GitHub, lowered the entry cost from $50,000 to a $3,000 student build. In agriculture, quadrupeds are emerging for crop phenotyping (Unitree A1 and Spot in cornfield trials), orchard inspection, and autonomous sheep mustering in Australia.
| Industry | Primary platforms | Typical tasks |
|---|---|---|
| Oil and gas | Spot, ANYmal, ANYmal X | Gauge reading, gas leak detection, thermography |
| Power generation | Spot, ANYmal, X30 | Substation inspection, transformer thermography |
| Construction | Spot, ANYmal | 3D scanning, BIM comparison, safety walkthrough |
| Mining | ANYmal, Vision 60 | Underground mapping, ventilation inspection |
| Military | Vision 60, V60-Q-UGV | Perimeter security, EOD, reconnaissance |
| Public safety | Spot, Vision 60 | SWAT recon, HazMat assessment |
| Search and rescue | Spot, ANYmal, X30 | Collapsed-building entry, victim location |
| Entertainment | Spot, Go1, Go2 | Live performance, advertising |
| Research | Mini Cheetah, Doggo, Go1 EDU | Algorithm development, education |
| Agriculture | Go1, Spot, A1 | Crop phenotyping, livestock monitoring |
| Healthcare | Spot, ANYmal | Hospital remote-presence, telemedicine pilots |
| Year | Milestone | Affiliation |
|---|---|---|
| 1968 | GE Walking Truck demonstrates teleoperated quadruped locomotion | General Electric, U.S. Army |
| 1980 | Marc Raibert founds CMU Leg Lab | Carnegie Mellon University |
| 1986 | MIT Leg Lab established; Quadruped trots dynamically | MIT |
| 1992 | Boston Dynamics founded | Marc Raibert |
| 2005 | BigDog unveiled; viral kick-recovery video | Boston Dynamics, DARPA |
| 2009 | MIT Cheetah introduced | MIT Biomimetic Lab |
| 2012 | LS3 AlphaDog enters Marine field testing | Boston Dynamics, USMC |
| 2016 | ANYbotics spun out of ETH Zurich | ANYbotics |
| 2017 | ANYmal wins ARGOS Challenge | ETH Zurich, ANYbotics |
| 2019 | Hwangbo et al. publish RL sim-to-real for ANYmal in Science Robotics | ETH Zurich |
| 2019 | Mini Cheetah backflip | MIT |
| 2020 | Spot commercial launch at $74,500 | Boston Dynamics |
| 2020 | Lee et al. publish blind RL locomotion in Science Robotics | ETH Zurich |
| 2021 | Unitree Go1 hits $2,700 | Unitree |
| 2022 | Miki et al. publish perceptive RL controller in Science Robotics | ETH Zurich |
| 2023 | Unitree Go2 priced from $1,600 | Unitree |
| 2024 | NVIDIA announces Project GR00T at GTC | NVIDIA |
| 2025 | Isaac GR00T N1 released as open foundation model | NVIDIA |
The global quadruped robot market was valued at roughly $2.0 billion in 2024 and is forecast to reach $7.1 billion by 2032 to 2033 at a 17 to 19 percent CAGR. Defense applications accounted for an estimated 48 percent of use cases in 2023, with industrial inspection and remote monitoring second.[14]
Boston Dynamics (Waltham, MA; acquired by Hyundai Motor Group in 2021 at a $1.1 billion valuation) holds the premium and brand-leadership position. Spot is the most recognizable industrial quadruped, with over 1,500 units shipped globally by 2025.
ANYbotics (Zurich, Switzerland) dominates high-end industrial inspection, especially in Europe and the Middle East. Its ATEX-certified ANYmal X is unique, and ANYmal D is widely deployed by Shell, BASF, OMV, and other process operators.
Unitree Robotics (Hangzhou, China) leads the consumer and education segment by units shipped, with reports of over 50,000 quadrupeds sold cumulatively by 2024.
Deep Robotics (Hangzhou, China) competes in mid-market industrial inspection with the X20 and X30 lines and pioneered the wheeled-legged form factor with the Lynx M20.
Ghost Robotics (Philadelphia, PA) is the dominant Western defense supplier, with Vision 60 deployed across the U.S. Air Force, DHS, U.S. Army, and dozens of allied militaries.
Xiaomi (Beijing, China) operates in the consumer and prosumer segments with CyberDog and CyberDog 2 at sub-$2,000 price points. Other notable manufacturers include MAB Robotics (Poland), Kawasaki Heavy Industries (Bex prototype, 2022), LEJU Robotics, and Sony's AIBO entertainment quadruped.
Quadrupeds have been the testbed for nearly every major locomotion algorithm later deployed on bipedal and humanoid robots. The Hwangbo 2019 RL recipe was extended to Cassie (Agility Robotics' bipedal robot from Oregon State University) and helped set a Guinness World Record for the fastest 100 meters by a bipedal robot in 2022. The same training infrastructure now drives Digit (Agility Robotics' humanoid), Tesla Optimus, Figure 02, and Boston Dynamics' all-electric Atlas. Quadrupeds remain a strictly easier locomotion problem (more contact points, lower center of mass) and a standard validation platform before tackling biped control.
Conversely, the abundance of quadruped data and platforms has made them the canonical testbed for foundation-model research outside manipulation. Quadrupeds will likely remain central to robotics research even as humanoids attract media attention, because they are cheaper, safer, and more reliable for outdoor, multi-hour experiments.
Quadrupeds still struggle with several problems. Battery life is typically 60 to 120 minutes unloaded and as short as 30 minutes under heavy compute or payload, limiting continuous deployment. Manipulation with a back-mounted arm is improving but remains well behind dedicated industrial manipulators. Stair-climbing in cluttered or unfamiliar environments is unreliable, and spiral staircases remain unsolved. Cost of high-end industrial platforms (over $100,000) limits adoption in cost-sensitive markets. Extreme weather operation requires expensive ingress protection generally restricted to flagship products.
Open research challenges include long-horizon autonomous task execution, robust LLM-grounded behavior trees, social acceptance in public spaces (especially for armed variants), and fleet-scale repair and maintenance protocols.