How four-legged robots went from MIT's Cheetah lab to industrial inspection: the actuator and control ideas in the papers, and what buyers now check.
A quadruped robot is a legged machine with four limbs, often called a robot dog. Legs let a robot choose where its feet land. The ETH Zurich team behind the ANYmal robot noted in 2019 that this lets legged robots overcome obstacles about as high as their legs are long, which wheels and tracks struggle to do on stairs, grated walkways or rubble.
Balance is what makes it hard. In a fast gait a quadruped spends much of each stride with one or two feet on the ground, or none. As the Cheetah 3 control paper puts it, during bounding or galloping the body is always underactuated, and every foot force has to stay inside a friction cone or the foot slips. The MIT Cheetah programme, in the MIT Biomimetic Robotics Lab, worked on both: actuators that produce and measure force quickly, and controllers that plan those forces in time. Quadrupeds are one branch of what is now called physical AI.
Earlier dynamic quadrupeds such as Boston Dynamics' BigDog and IIT's HyQ used hydraulic actuators. The 2019 ETH paper notes the drawbacks: such robots are hard to build below about 40 kg, and their smoke and noise kept them outdoors.
BigDog also used a lot of energy for its weight and speed. In a 2015 journal paper, Sangok Seok, Kim and colleagues compared robots by total cost of transport: power drawn, divided by weight times speed. Their chart puts BigDog at 15 and ASIMO at 2. The MIT Cheetah, a 33 kg robot trotting at 6 m/s (22 km/h), drew 973 watts from its battery for a figure of 0.5, which the authors say rivals running animals of the same size. They credit four principles: motors with high torque density, electronics that recover energy, a low-loss transmission and light legs.
A 2017 paper in IEEE Transactions on Robotics, led by Patrick Wensing, named the approach proprioceptive actuation. A motor behind a low gear ratio can be back-driven, so an impact at the foot can turn the motor backwards and the gears are spared much of the shock. The same property lets the robot estimate the force at the foot from the current in its motors. In bounding, the leg controlled contact forces with ground contact as short as 85 milliseconds and peaks above 450 newtons.
The same paper sets out the trade-off: a low gear ratio reduces the inertia reflected through the gearbox, but producing torque that way generates more heat in the motor. A slow robot carrying loads, the authors write, could gain on energy from a higher ratio compared with an agile robot built to run. With this hardware the MIT Cheetah 2 bounded at 6 m/s and jumped obstacles on its own, though it could move only in the forward and vertical plane.
The Cheetah 3 paper, by Gerardo Bledt and colleagues at IROS 2018, describes a 45 kg robot with nearly identical actuators on all three joints of each leg, so it could control foot forces in three dimensions. Each actuator pairs a custom high-torque motor with a single-stage 7.67:1 planetary gearbox. That is slightly higher than Cheetah 2's 5.8:1, a choice made to improve load-carrying and low-speed efficiency. The four legs together weigh 2.7 kg, 6% of the robot's mass, and ground forces are controlled with no force sensors, torque sensors or springs at the joints or feet.
The paper reports a lowest cost of transport of 0.45, in a trot, and a blind climb up a flight of stairs. With no external sensors, the robot assumed the world was flat and found each step by detecting unexpected contact with its feet. It then adjusted its swing trajectories and posture and reached the top of a test staircase with a 17.8 cm rise.
The controller came from a companion paper by Jared Di Carlo and colleagues. It simplifies the robot's dynamics so that force planning becomes a convex problem, a form a computer can solve quickly and reliably. The robot planned up to half a second ahead, solved each plan to optimality in under a millisecond, and replanned 20 to 30 times a second. With one set of gains it ran gaits from a trot to a full 3D gallop, at forward speeds up to 3 m/s. The shortcut holds because of the hardware: Katz's 2018 thesis notes the legs are light enough that most controllers on the MIT Cheetah ignore their mass and inertia.
Benjamin Katz's 2018 MIT thesis, supervised by Kim, starts from a practical problem: a robot the size of the MIT Cheetah is costly, and risky to test new controllers on. Katz set out to build a cheap actuator for a smaller research robot. It uses mass-produced brushless motors made for radio-controlled drones and aircraft, a 6:1 single-stage planetary gearbox and a built-in motor controller. According to the thesis, the hardware for the whole quadruped cost less than one actuator of the MIT Cheetah.
The robot weighed 9 kg with its battery and performed a full 360 degree backflip from flat ground, which the author believed was a first for a quadruped. Di Carlo's 2020 thesis records 11 Mini Cheetahs built in the MIT lab for distribution to research groups, with the open-source control software already used by seven other groups around the world.
Mini Cheetah's first controller reached 2.45 m/s, according to a follow-up paper from the lab, and that paper pushed it to 3.7 m/s by combining model predictive control with whole-body control. In 2022 MIT's Improbable AI Lab trained a neural network controller in simulation that sustained speeds up to 3.9 m/s and ran on grass, ice and gravel.
The ETH Zurich and Intel study by Jemin Hwangbo and colleagues, published in Science Robotics in 2019, took that route earlier. They trained a neural network policy in simulation and transferred it to ANYmal, a medium-dog-sized quadruped. On the real robot it reached 1.5 m/s, against the platform's earlier record of 1.2 m/s, and recovered from falls in complex positions.
Among robots with electric actuators, the ETH paper calls the MIT Cheetah "one of the most promising legged systems of this kind", while noting that it was a research platform optimised mainly for speed and not thoroughly evaluated for battery life, turning, mechanical robustness or outdoor use. The authors write that ANYmal was designed to be robust, reliable and versatile.
The MIT papers reported top speeds of 3 to 6 m/s. Spot's listed maximum is 1.6 m/s, and ANYmal quotes a normal walking speed of 0.75 m/s. Each of these pages gives a dust and water rating or a hazardous-area approval next to its speed figure. Battery life and outdoor use are two of the measures the ETH authors listed in 2019 as not thoroughly evaluated on the MIT Cheetah.
For a site considering one of these robots, the deciding facts sit on the same sheets: the hazardous-area zone it is approved for, whether one charge covers the inspection route or docking stations are needed along it, and whether its sensors can read the gauges, temperatures and leaks that matter on that site.
A four-legged robot that balances by controlling the forces its feet apply to the ground. Smaller commercial models are often called robot dogs.
Yes. Boston Dynamics, ANYbotics and Unitree each market quadrupeds for inspection, and ANYbotics offers a version approved for Zone 1 explosive atmospheres in oil, gas and chemical plants.
Their makers say they can run routine inspection rounds, including stairs and grated walkways, recording visual, thermal and acoustic data, so fewer people need to enter hazardous areas. ANYbotics also lists less manual work and no facility shutdowns for inspections.
The summit lists him as Prof. Sangbae Kim, Robotics Architect, Meta & MIT. He is the senior author of the MIT Cheetah papers cited here, and he supervised the MIT theses behind Mini Cheetah and its open-source control software.
Physical AI and robotics is one of the four keynote-led panels at the Davos Super Intelligence Summit 2027, on 19 January 2027 at the Grandhotel Belvedere in Davos. Speakers announced so far include Prof. Sangbae Kim, Robotics Architect, Meta & MIT, and J. Bruce Chan, Managing Director, Transportation and Robotics, Stifel Financial Corp. Access is by invitation only; the four panels are set out in the programme.