One Giant Leap for Robotics

Legged robots may be impressive feats of engineering, but they still lag far behind the agility and dexterity of humans and animals. One of the primary reasons for this gap is the complexity of replicating the intricate biomechanics and neural control systems that are responsible for natural movement. Humans and animals possess highly adaptable musculoskeletal structures coupled with sophisticated nervous systems that enable them to navigate varied terrains, adjust to unexpected obstacles, and maintain balance effortlessly. In contrast, legged robots often rely on pre-programmed algorithms or simple feedback control mechanisms, which lack the versatility and adaptability of biological systems.

There are many efforts underway to close this gap, because the potential utility of more agile robots is huge. In search and rescue missions, for instance, robots with human-like agility could navigate disaster zones more effectively, accessing otherwise inaccessible areas to locate survivors or assess the extent of damage. Similarly, in exploration scenarios, such as planetary exploration or mapping of hazardous environments like nuclear reactors or deep-sea trenches, agile robots could maneuver through complex terrain with ease, gathering data and performing tasks efficiently. These nimble robots could also be a huge boon to industrial applications.

We previously reported on ETH Zurich’s quadrupedal ANYmal robotics platform (see here and here) as the team has continued to enhance its capabilities. In their latest work, they have elevated ANYmal’s agility to a new level. Their goal was to design a robot that can complete a parkour course, which can even be quite challenging for humans. In order to achieve that goal, they took an innovative approach that solves many persistent issues that plague existing systems.

These issues include the need for adaptability in motion planning due to unpredictable terrains and obstacles, difficulties in accurate state estimation caused by high-impact forces, limitations in environmental perception due to self-occlusions and sensor field of view, the complexity of motion planning considering the robot's capabilities and controller limitations, and the critical importance of minimizing latency during fast motions to avoid catastrophic outcomes.

The researchers’ solution to these problems has three main components: perception, locomotion, and navigation. The perception module uses data from onboard cameras and LiDAR to understand the terrain and surroundings with the assistance of a neural network, while the locomotion module contains various skills for traversing different obstacles. A particular skill is chosen by the navigation module based on its likelihood to succeed, as determined by previous results. The navigation module also provides the commands needed to carry out the chosen skill.

A number of experiments were conducted to determine how well ANYmal could deal with challenging obstacles that lie in its path. It was found that the robot could move at up to two meters per second while autonomously making decisions that helped it to reach a target destination. The system’s navigation capabilities proved to be valuable in ducking under obstacles, squeezing through narrow passes, and climbing over tall barriers. Real-time operation of the integrated components enabled ANYmal to react to unexpected conditions and changing grips without losing its footing or failing to meet its main objectives.

These upgrades may not have made ANYmal more capable than humans — or even as capable as humans — however, the agility demonstrated exceeds what is possible with virtually any other robot available today. And that might mean that these robots could take over a number of dangerous jobs that only humans can do today.