Legged robots: machines that walk over rough ground

A legged robot moves by placing its feet on the ground one after another, much as an animal does, instead of rolling on wheels or tracks. Four legs let it step over obstacles and cross loose or uneven ground that can stop a wheeled machine. That matters for industrial sites such as offshore platforms, where routine inspection rounds are carried out on foot. This article covers four-legged machines and the control problem behind them. Two-legged and human-shaped machines have their own article.

How it works

The body and legs are the easy part to picture. The hard part is balance and footing. The control software has to decide when each foot lands, where it lands and how the body shifts, many times a second, while the ground keeps changing underfoot.

Researchers have trained this software in computer simulation using reinforcement learning, then moved it onto the physical robot. This is called sim-to-real transfer. A 2019 Science Robotics paper by Hwangbo and colleagues describes a learned locomotion policy transferred to the ANYmal quadruped (arXiv abstract). A 2020 Science Robotics paper by Lee and colleagues uses a neural network that acts on proprioceptive signals, meaning measurements from the robot's own body sensors such as joint positions and body motion (arXiv abstract).

What has been demonstrated

These are peer-reviewed results we could confirm, with the conditions the sources give.

• Rough terrain (Science Robotics, 2020). Lee and colleagues report that their controller, trained in simulation, stayed robust on mud and snow, rubble, thick vegetation and gushing water. These conditions were not seen in training, and the authors describe the controller as running on two generations of ANYmal (arXiv abstract). The abstract gives no trial counts, distances or timings, so the scale of these tests is not known from it.
• Speed and falls (Science Robotics, 2019). Hwangbo and colleagues state that ANYmal followed body velocity commands precisely and energy-efficiently, ran faster than before and recovered from falls, including in complex configurations (arXiv abstract). The abstract gives no numerical speeds or recovery rates.
• A marathon on one charge (Nature, 2026). Lee and colleagues report that the RAIBO2 quadruped completed a full marathon in 4 hours, 19 minutes and 52 seconds on a single battery charge, with a total cost of transport of 0.25, against a human benchmark of 0.37 they cite. Cost of transport measures the energy used to move, allowing for the robot's weight; lower is better. They report more than three times the travel range per charge of existing quadrupeds (Nature). A Nature Portfolio press release dated 24 September 2026 names the race as the 22nd Sangju Marathon in South Korea, run alongside human runners, and says the robot used 1,280 Wh of a 2,016 Wh battery (Nature Asia). We have read the abstract and the press release, not the full paper, so the detailed test conditions are not covered here. This is a single run on one course.
• Energy use (review, 2018). A review in Frontiers in Robotics and AI compares specific resistance, a measure closely related to cost of transport, where lower is better. It gives 0.2 for a horse trotting or a human walking, 15 for the hydraulic BigDog quadruped and 0.5 for the MIT Cheetah. The authors conclude that the energy economy of today's robots "lags far behind" that of biological systems (Frontiers). The review is from 2018, and the RAIBO2 result above shows that newer machines can differ.

What is reported but not independently checked

These figures come from manufacturers. We did not test them.

• Boston Dynamics Spot. The company's product page lists a weight of 33.8 kg with battery, a top speed of 1.6 m/s, slopes of up to 30 degrees, steps of up to 300 mm and an IP54 dust and splash rating. It lists a 564 Wh battery, an average runtime of 90 minutes, a 60 minute recharge and a maximum payload of 14 kg. It also notes that runtime "may vary depending on payloads and environmental factors" (Boston Dynamics). Specifications were read on 23 October 2026 and may change.
• Offshore inspection (company case study, 2021). ANYbotics reports that PETRONAS carried out a multi-week installation of ANYmal C on its Dulang B offshore platform, after first field-testing the robot on onshore training platforms. The article is dated 22 January 2021, gives no year for the offshore deployment and reports no quantified inspection or cost results (ANYbotics). ANYbotics makes the robot, so treat this as a company account.

Limits and open problems

• Battery life. Spot's listed average runtime is about 90 minutes, and the maker says this varies with payload and environment. The RAIBO2 press release says quadruped robots have been limited to a range of about 20 km by battery life (Nature Asia).
• Energy cost. The 2018 review found that legged machines used far more energy per distance travelled than animals (Frontiers). RAIBO2 is reported to beat the human figure, but it is one research robot, not a commercial unit.
• Payload. Spot's listed maximum of 14 kg limits the sensors and tools one robot can carry on a mission. We found no reliable figures for arm or gripper work on quadrupeds.
• Evidence. Most industrial evidence comes from the vendors. The PETRONAS case study gives no independent return-on-investment or inspection figures. The learned-locomotion results are published in peer-reviewed journals, but we have summarised them from the abstracts only.
• Hazardous sites. In January 2021, ANYbotics said an Ex-certified ANYmal, designed to the IECEx standard for Zone 1 areas, was being co-developed with PETRONAS. It said it would offer the version globally, aimed for commercial use in 2022 and had begun onboarding early customers (ANYbotics). Those are company statements from 2021, and we have not checked whether the target was met. Certification matters because it decides whether these robots can enter places with explosive gas.
• Gaps. We did not find sourced figures for purchase price, noise, safety around people or failure rates. Buyers will need all of these before wide use, and this article cannot answer them.

Speculative uses

These have not been demonstrated in the sources we checked.

• Inspecting a site after a gas leak or structural failure, where people cannot safely enter.
• Night patrols of unstaffed plants or substations on fixed routes, with automatic reports.
• Regular scans of building sites to compare progress against plans.

What to watch next

• Independent comment on, or repeats of, the RAIBO2 marathon result on other courses and conditions.
• Independent trials of runtime and payload, rather than manufacturer specification sheets.
• Certified explosion-proof versions for hazardous areas, and whether early customers publish results.
• Published energy figures for commercial units, so cost of transport can be compared over time.
• Noise and safety data for use near workers, which the sources we checked do not cover.