Scientists have successfully utilized artificial intelligence (AI) to train a four-legged robotic creature, resembling a dog, to conquer a basic parkour course with a speed of up to 5 mph (8 km/h).
In newly released footage, the quadruped robot, named “ANYmal,” demonstrates its abilities by scaling a 3-foot-tall (1 meter) wooden crate, leaping across a gap of the same distance between two large crates, and safely descending from one of them.
The machine has acquired various maneuvers, including walking, crouching, climbing, and jumping, which researchers believe could prove invaluable in enabling it to navigate physical obstacles during search and rescue operations.
Dubbed “robot dogs” due to their backward-facing knees that resemble those of canines, robots like ANYmal are becoming increasingly prominent, as stated by the Institute of Electrical and Electronics Engineers (IEEE).
The study outlining their training techniques was recently published on March 13 in the journal Science Robotics.
“By aiming to match the agility of free runners, we can better understand the limitations of each component in the pipeline from perception to actuation, circumvent those limits, and generally increase the capabilities of our robots,” explained the scientists in a statement. They further emphasized how this advancement could open avenues for numerous applications, including search and rescue missions in complex terrains.
While acknowledging that most robot dogs, such as Boston Dynamics’ Spot, lack such agility, the scientists highlighted how neural networks, structured similarly to neurons in the human brain, were utilized to train ANYmal, enhancing its perception, locomotion, and navigation skills.
The team designed the “module pipeline” for AI training based on the movement patterns of free runners, adapting them for a quadrupedal form.
Following training in simulated parkour courses, ANYmal successfully traversed a real course, as depicted in the video footage. Additionally, the robot exhibited nimble maneuvering capabilities, navigating effortlessly under tables.
This learning pipeline can now be implemented in other dog-like robots equipped with necessary hardware, including sensors and actuators, to enhance their terrain navigation abilities.
While the robot has not yet undergone testing in simulated disaster environments, the scientists intend to further refine the learning pipeline to enhance the agility of such robots.
