Legged robots

Also known as Quadruped robots, Quadrupedal robots, Four-legged robots, Robot dogs

Four-legged robots walk over stairs, mud, rubble and other rough ground where wheels struggle, and makers now sell them for inspection work at industrial sites. Learned control software has made them more robust on uneven terrain, but Spot's battery life of about 90 minutes and payload of about 14 kg still limit what a robot can do on a mission.

CommercialDemonstratedReviewed 22 Oct 2026
AI illustration · Illustrative scene, not a photograph of the specific device or experiment.

In plain language

A legged robot moves by placing its feet on the ground in turn, much as an animal does. Four legs let it step over gaps, climb steps and cross loose ground that would stop a wheeled machine. The two best-known commercial examples are Boston Dynamics' Spot and ANYbotics' ANYmal. Both are marketed mainly for inspection work, such as reading gauges and taking thermal or sound readings at industrial sites.

The hard part is balance and footing. Researchers now train the control software in computer simulation using reinforcement learning, then move it onto the real robot. The resulting controllers can keep walking on mud, snow, rubble and thick vegetation in tests. Even so, Spot runs for about an hour and a half on one battery charge, and it can carry only a limited load.

Technical explanation

Many recent quadruped controllers are neural network policies. A 2019 Science Robotics paper by Hwangbo and colleagues describes a learned locomotion policy transferred to the ANYmal quadruped. The authors report that the robot follows high-level body velocity commands precisely and energy-efficiently, runs faster than before, and recovers from falling even in complex configurations. The abstract gives no numerical speeds or recovery rates. A 2020 Science Robotics paper by Lee and colleagues (arXiv preprint 2010.11251) describes a controller that acts on proprioceptive signals, which are measurements from the robot's own body sensors such as joint encoders and an inertial measurement unit. The authors describe zero-shot generalisation from simulation to natural environments, and report that the controller stays robust on mud and snow, rubble, thick vegetation and gushing water, conditions not seen in training.

Makers publish specifications for commercial units. Boston Dynamics lists Spot at 33.8 kg with its battery, a maximum speed of 1.6 m/s, a maximum slope of plus or minus 30 degrees, a maximum step height of 300 mm and an IP54 ingress rating. It gives a 564 Wh battery, an average runtime of 90 minutes (standby 180 minutes), a recharge time of 60 minutes, a maximum payload of 14 kg, and an operating range of -20 to 55 degrees Celsius. The maker notes that runtime may vary with payloads and environmental factors. These are manufacturer figures and have not been independently tested in this profile.

Energy use is the central engineering cost. A 2018 review in Frontiers in Robotics and AI tabulates specific resistance, which it equates with cost of transport, for several machines. Values include 0.2 for a horse trotting or a human walking, 15 for the hydraulic BigDog quadruped, and 0.5 for the MIT Cheetah. The review concludes that the energy economy of today's robots lags far behind that of biological systems. A Nature press release (2026) reports that the RAIBO2 quadruped completed a marathon on one battery charge with an average cost of transport of 0.25, against a stated human value of 0.37. The paper itself was not reviewed here. In industry, ANYbotics reports that PETRONAS carried out a multi-week installation of ANYmal C on its Dulang B offshore platform, in a case study dated January 2021. The same article says an Ex-certified version designed to the IECEx standard for Zone 1 areas is in co-development.

What the evidence says

Each statement is typed by the strength of its evidence and linked to the source that supports it.

Demonstrated Shown in a published experiment, test or shipping product.

Lee and colleagues trained a neural network controller in simulation, and deployed it on the ANYmal quadruped. The controller acts on a stream of proprioceptive signals, which are measurements from the robot's own body sensors.

Sources: [1] · Abstract

The authors report that the Lee et al. controller stays robust on mud and snow, rubble, thick vegetation and gushing water, conditions not seen in training, and describe zero-shot generalisation from simulation to natural environments.

Sources: [1] · Abstract

Hwangbo and colleagues transferred a learned locomotion policy to the ANYmal robot. The authors state that it followed high-level body velocity commands precisely and energy-efficiently, ran faster than previous methods, and recovered from falling even in complex configurations. The abstract gives no numerical speeds or recovery rates.

Sources: [2] · Abstract

A 2018 review tabulates specific resistance (equivalent to cost of transport) of 0.2 for a horse trotting or a human walking, 15 for the hydraulic BigDog quadruped and 0.5 for the MIT Cheetah, and concludes that the energy economy of today's robots lags far behind that of biological systems.

Sources: [6] · Abstract and comparison table · medium confidence

Reported, unverified Stated by a company or team but not independently confirmed.

Boston Dynamics lists Spot's maximum payload as 14 kg, its battery capacity as 564 Wh, and its average runtime as 90 minutes, and says runtime may vary depending on payloads and environmental factors.

Sources: [3] · Specifications section · medium confidence

Boston Dynamics lists Spot's maximum speed as 1.6 m/s, maximum slope as plus or minus 30 degrees, maximum step height as 300 mm, and ingress protection as IP54.

Sources: [3] · Specifications section · medium confidence

According to ANYbotics, PETRONAS carried out a multi-week installation of ANYmal C on its Dulang B offshore platform. The article, dated 22 January 2021, does not give the year of the deployment and reports no quantified inspection results.

Sources: [4] · Article body · medium confidence

A Nature press release reports that the RAIBO2 quadruped completed a marathon on a single battery charge with an average cost of transport of 0.25, against a stated human value of 0.37, and states that quadruped robots have been limited to a range of about 20 km by battery life.

Sources: [5] · Press release body · medium confidence

Limitations and open problems

  • Battery life is a major constraint: Spot's listed average runtime is about 90 minutes, and Boston Dynamics says this varies with payload and environment. The Nature press release on RAIBO2 says quadruped robots have been limited to a range of about 20 km by battery life.
  • Legged locomotion is energy-hungry. The 2018 Frontiers review states that the energy economy of today's robots lags far behind that of biological systems.
  • Payload is limited. Spot's listed maximum payload is 14 kg, which restricts the sensors and tools a robot can carry on a single mission.
  • Most industrial evidence is vendor-reported. The sources reviewed here, including the PETRONAS case study, give no independent return-on-investment figures and no quantified inspection metrics.
  • No reliable figures were found in the sources reviewed for purchase cost, or for arm and gripper manipulation on quadrupeds, so this profile does not state them.

Possible future uses

Speculative

Ideas, not predictions. None of these has been demonstrated.

  • Inspecting sites after an incident, such as a gas leak or structural failure, where it is unsafe for people to enter.
  • Scheduled night-time patrols of unmanned plants or substations, with fixed routes and automatic reporting.
  • Scanning construction sites on a regular schedule to compare progress against plans.

Timeline

  1. 16 Jan 2019Hwangbo and colleagues publish a method for training a neural network locomotion policy and running it on the ANYmal quadruped in Science Robotics. Sources: [2]
  2. Oct 2020Lee and colleagues publish a controller that acts on proprioceptive signals and which the authors report walks ANYmal over mud, snow, rubble and thick vegetation, in Science Robotics. Sources: [1]
  3. 22 Jan 2021ANYbotics publishes a case study on a multi-week installation of ANYmal C on PETRONAS's Dulang B platform. Sources: [4]
  4. 23 Sept 2026A Nature paper reports the RAIBO2 quadruped completing a marathon on one battery charge, as described in a Nature press release. Sources: [5]

Futures this could help bring about

Glossary

Proprioception
Sensing of the robot's own body state, such as joint positions and body motion, as opposed to sensing the outside world with cameras or lidar.
Sim-to-real transfer
Training a control policy in a computer simulation and then running it on physical hardware, usually with some adjustment.
Cost of transport
A measure of energy used per unit weight and per unit distance travelled; lower values mean a more energy-efficient gait.
Ingress protection (IP) rating
A standard code showing how well a device resists dust and water; IP54 means protected against dust and splashing water.

Sources

  1. peer reviewed paper
    Learning quadrupedal locomotion over challenging terrain
    Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen et al.. Science Robotics, Oct 2020. DOI: 10.1126/scirobotics.abc5986. Accessed 22 Oct 2026.
  2. peer reviewed paper
    Learning agile and dynamic motor skills for legged robots
    Jemin Hwangbo, Joonho Lee, Alexey Dosovitskiy et al.. Science Robotics, 16 Jan 2019. DOI: 10.1126/scirobotics.aau5872. Accessed 22 Oct 2026.
  3. company announcement
    Spot (Boston Dynamics product page)
    Boston Dynamics, 2026. Accessed 22 Oct 2026.
  4. company announcement
    ANYmal C inspection of PETRONAS Dulang B platform (ANYbotics article)
    ANYbotics, 22 Jan 2021. Accessed 22 Oct 2026.
  5. news
    Robotics: Robot runs a marathon (Nature)
    Nature Asia (press release reprinting Nature), 24 Sept 2026. Accessed 22 Oct 2026.
  6. review article
    An Overview on Principles for Energy Efficient Robot Locomotion
    Navvab Kashiri, Andy Abate, Sabrina J. Abram et al.. Frontiers in Robotics and AI, 11 Dec 2018. DOI: 10.3389/frobt.2018.00129. Accessed 22 Oct 2026.

Revision history

3 changes since this profile was created.

Changes on 22 Oct 2026
Published on owner instruction (content loop iteration 62)0bb7456
Sources and claims re-checked against originals409337d
Initial profile drafted from primary sourceseaac9eb