When this build is up and running, it moves in two ways at once: the four legs can step while the motorized wheels spin underneath them. The result is a red quadruped called ANYmal that can drive along flat ground at up to 4 meters per second and still climb over obstacles, having navigated grassy hills, indoor stairs, and paved roads. It can even tackle challenging terrain blindly, which sounds remarkable for a machine with this kind of hybrid gait.
The construction is worth a close look. The central red body houses the electronics and sensors, and each leg ends in a wheel with a silver rim and a black tire. The legs run on series elastic actuators while the wheels use direct drive. On the software side, a whole-body model predictive controller optimizes wheel and torso motions together, and it relies on a single parameter set for every behavior, so the total parameter count stays low. The optimization avoids hand-made heuristics, which means almost no tuning is needed. That seems like a smart call to me, since earlier approaches apparently needed gain tuning per gait.
For route planning, traversability comes from a learning-based cost map, essentially a grid of costs. During testing, researchers kept an eye on the machine with laptops and handheld remote controllers.
One open item: fall recovery has not been worked on yet, though the maker sees no reason it would not work. Another practical point is price. The prototype reportedly cost several hundred thousand dollars, but mass production would bring that down to a few tens of thousands. The maker also sees uses in search and rescue, industrial inspection, delivery, and maintenance if an arm is mounted.
Project by Robotic Systems Lab: Legged Robotics at ETH Zürich. Watch it on their YouTube channel: