ANYmal robot showcases remarkable parkour skills after undergoing extensive neural network training.

By emphasizing walking, crouching, climbing, and jumping, the redesigned learning approach equips ANYmal to adeptly maneuver through obstacles, potentially enhancing its effectiveness in future rescue operations.

(Photo : LUCAS BARIOULET/AFP via Getty Images)

An amateur performs a "Parkour" in the streets near Montpellier's Saint Roch Station, on the sidelines of the FIG Parkour World Cup taking place at the FISE World Series 2018 in Montpellier on May 12, 2018.

Excelling in Parkour Feats

A quadrupedal robot resembling a dog ANYmal has undergone significant enhancements to its agility through a novel framework, empowering it to traverse a simple parkour course at speeds of up to 6 feet per second. 

Parkour is an athletic activity centered around traversing obstacles in urban environments, continuing to attract a broad audience.

With its updated learning methodology prioritizing crawling, jumping, climbing, and crouching, the robot is poised to potentially navigate through physical barriers, such as crawling beneath and leaping over obstacles, during search and rescue missions.

Interesting Engineering reported that their latest method stands out for its ability to devise paths for complex scenarios without relying on expert demonstration, offline computation, prior environmental knowledge, or explicit consideration of contacts.

Currently, quadrupedal bio-inspired robots, exemplified by ANYmal, bear a noticeable lack of agility when compared to their animal counterparts. Their dynamic movements, varied points of contact, and limited perceptual sensor range pose challenges to their agile navigation.

Translating Parkour Maneuvers into Real-World Actions

Guided by virtual obstacle courses, this framework facilitated ANYmal in translating parkour maneuvers into real-world actions.

As elucidated by the team, this methodology involves instilling ANYmal with advanced locomotive capabilities tailored for a spectrum of challenges, encompassing walking, jumping, climbing, and crouching. 

These skills are meticulously curated and orchestrated throughout the environment via a high-level policy. Characterized by its hierarchical structure, the navigation policy adapts its behavior contextually and leverages the distinctive capabilities of each skill.

ANYmal successfully executed parkour maneuvers in a real-world setting with the assistance of the framework, honing its skills through practice on virtual obstacle courses. The robot showcased remarkable agility as it gracefully ascended, leapt over sizable obstacles, and swiftly navigated through confined spaces.

Also Read: Robots Are Being 'Trained' To Walk Using Virtual Obstacle Courses

In a bid to address this issue, the team devised an enhanced neural network approach comprising three distinct modules dedicated to sensing, locomotion, and navigation.

Drawing inspiration from the methodologies employed by parkour practitioners, the team crafted a module pipeline mirroring the process by which human parkour athletes assess and execute impressive feats of athleticism. 

As published in Science Robotics, ANYmal adeptly maneuvered through diverse configurations of randomly positioned obstacles to reach its target. 

Despite being trained exclusively on simulated data, the real-world trials demonstrated seamless adaptation onto hardware, underscoring the robot's capacity to surmount consecutive obstacles at speeds of up to 6 feet per second.

The team emphasized the importance of matching the agility of free runners, enabling a comprehensive understanding of the limitations inherent in each component of the pipeline, from perception to actuation. 

By addressing these constraints, the team anticipates an expansion of the robot's capabilities, opening avenues for diverse applications such as search and rescue missions in complex environments like collapsed structures or rugged terrains.

Related Article: [VIDEO] Boston Dynamics' Parkour Robot CRUSHES Obstacle Course