Each leg typically has three motorised joints, and a controller running hundreds of times per second decides where to put each foot to keep the body upright. Sensors in the joints measure torque so the robot can feel ground contact, and an inertial measurement unit tells it which way is up.
Modern quadrupeds increasingly use controllers trained by reinforcement learning in simulation, which makes them robust to slips, pushes and soft ground without hand-tuned rules for every situation.
Commercial examples such as Boston Dynamics Spot and ANYbotics ANYmal carry cameras, thermal sensors and gas detectors on routine inspection routes through substations, oil platforms and factories, places built for human feet rather than for wheels.
