Imma Perfetto
Cosmos science journalist
Researchers have developed a new way to train a 4-legged robot to clamber its way across a parkour obstacle course. Experts say the capability puts its designers at the forefront of robot technology.
The acrobatic AT-AT could reach speeds of up to 2m/s while navigating the route completely on the fly.
“Acrobatic skills such as parkour have been used to evaluate capabilities of robot hardware and algorithms for a number of years now,” Professor Ian Manchester, director of the Australian Centre for Robotics at the University of Sydney, told Cosmos. Manchester was not involved in the research.
“What was new in this approach was that the robot could perform these feats on a course that it hadn’t seen before. Many previous methods were essentially choreographed for a specific sequence of motions.”
To achieve this, the team from the Robotic Systems Lab at the ETH Zurich gave the robot the ability to track its rapidly changing surroundings and select a path forward using a sequence of movements it deemed most likely to succeed.
The feat has been described in a new study published in the journal Science Robotics.
“For several years now this group at ETH Zurich has been one of a handful of teams across industry and academia that are on the forefront of this technology,” says Manchester.
“This paper draws on a lot of research from many different groups, and probably the most interesting thing is the way they put it all together to get a working system this impressive.”
Three learning modules used 1 or more neural networks to achieve this.
“The perception module receives point cloud measurements from the onboard cameras and the LiDAR (light detection and ranging) and computes an estimate of the terrain around the robot,” the authors wrote in the study.
“The locomotion module contains a catalogue of locomotion skills that can overcome specific terrains.” This includes jumping, climbing up and down, crouching, and walking.
Finally, the navigation module “guides the locomotion module in the environment by selecting which skill to activate and providing intermediate commands.”
These modules were trained in simulation, but experiments showed they were more than capable of pulling off parkour in the real world too.
The perception pipeline correctly reconstructed the scene despite the robot’s fast speeds, while the locomotion module “performed precise and agile movements, sometimes on narrow boxes barely the size of the robot’s footprint and leveraged the system’s full range of motion to pass higher obstacles.”
Lastly, the navigation module “used the available information and its intrinsic knowledge of each skill’s capabilities to guide the robot around the course on a feasible path.”
Last week, Cosmos reported that the same lab group had trained a 4-legged robot to play badminton against human opponents. This week, the challenge was parkour.
It makes for incredible footage, but why train robots to tackle human sports?
“Similar to sporting events for humans, these sorts of challenges are used to test robots mainly because they are hard, not because they are directly linked to applications,” says Manchester.
“But they provide a benchmark and a motivation that pushes the development of robot capabilities.”
These capabilities may include navigating difficult and unstructured terrain for applications where time is of the essence, such as search and rescue in collapsed buildings or complex natural settings.
“These capabilities will eventually be beneficial in applications such as manufacturing or assistive robots for the elderly or disabled, which may be less visually exciting but will ultimately be more impactful,” says Manchester.
