ANYmal has been successfully navigating the stony terrain of Swiss hiking trails for some time. Researchers at ETH Zurich have now expanded its skill set, teaching the quadrupedal robot to excel in parkour, a popular sport that involves smoothly overcoming obstacles in urban environments. ANYmal has also proven its proficiency in handling challenging terrains typically found on construction sites or in disaster zones.
To equip ANYmal with these new abilities, two teams within ETH Professor Marco Hutter’s group at the Department of Mechanical and Process Engineering took different approaches.
Exploring mechanical options
One team, including ETH doctoral student Nikita Rudin, who practices parkour in his spare time, focused on pushing the boundaries of what legged robots like ANYmal can achieve. By utilizing machine learning, Rudin successfully taught ANYmal to scale obstacles and execute dynamic manoeuvres to navigate them.
ANYmal learned through trial and error, using its camera and artificial neural network to identify obstacles and determine the best course of action based on previous training.
Rudin acknowledges the potential for further enhancements, such as enabling the robot to tackle complex terrains like disaster zones with scattered debris.
Integrating new and traditional technologies
Another project, led by ETH doctoral student Fabian Jenelten, aimed to prepare ANYmal for tasks in hazardous environments. Jenelten combined machine learning with model-based control, a proven method in control engineering, to teach the robot precise maneuvers, like navigating through gaps in rubble piles. This hybrid approach allows ANYmal to adapt movement patterns to unforeseen circumstances.
As a result, ANYmal has improved its stability on slippery surfaces and unstable terrain, making it suitable for tasks such as inspecting collapsed structures in disaster areas.
