ANYmal — the academic quadruped that became an inspection robot

ANYmal started life as a research robot at ETH Zurich's Robotic Systems Lab. It was the platform behind some of the earliest reinforcement-learning-trained locomotion papers — the kind where the robot learns to walk in simulation and then transfers to hardware without a single hand-tuned gain. ANYbotics commercialized it in 2017.

Why industry buys it

ANYmal's pitch is not "look at this cute robot." It's: 24/7 autonomous inspection of dangerous places. Refineries, offshore rigs, electrical substations. The robot does a pre-programmed route, takes thermal + visual readings of equipment, flags anomalies, and goes back to charge — all without a human nearby.

The IP67 rating means it can walk in rain, dust, and oily floors. The thermal camera reads the temperature of pipes and joints to within a degree. The robot has logged tens of thousands of hours in real industrial sites, mostly in Switzerland and the Middle East.

What makes ANYmal's locomotion good

The legs use series elastic actuators — springs in line with the motors. This makes the joints compliant (they give a little when something pushes back), which is critical for surfaces with unknown friction. The control loop is RL-based: a policy network trained in simulation outputs joint torques given proprioception inputs.

The team has published several Nature/Science papers on the locomotion architecture, including a notable one in 2022 showing recovery from any fall pose without explicit programming.

Price

Not advertised publicly. Industry rumours put a full ANYmal D system at around $200,000 USD — substantially more than a Spot, justified by the autonomous inspection use case where one robot replaces an inspection team.

The Atlas entry Reinforcement Learning covers the training approach used here.