A team of researchers from ArcLab at the University of Hong Kong has created a robot dog that can walk on two legs on a complex surface, maintain balance even during shocks, and pick itself up after a fall. About this informs Interesting Engineering.
The innovation of the device consists in the use of a bio-inspired “TumblerNet” controller, which works on the basis of deep reinforcement learning (Deep Reinforcement Learning). Thanks to it, the robot imitates human gait and can adapt to different types of surfaces — from sand to loose stones or soft mats.
“TumblerNet” combines data about the center of mass and the center of pressure into a single system that reproduces the balance inherent in humans. This allows the robot to freely transition from quadrupedal to bipedal mode, move forward, turn, and traverse curvilinear trajectories using both its forelimbs and hindlimbs.
One of the robot’s key advantages is its durability. It can withstand being pushed, bumped and even dropped without the need for an external stabilization system. Even if the robot catches on objects, it is able to restore its position on its own.
The development was also tested on the beach, one of the most difficult environments for movement. Despite the unstable surface, the robot successfully moved on two legs, demonstrating high adaptability and reliability.
Unlike traditional four-legged robot dogs, this model is better integrated into the human environment: it is able to climb stairs, overcome obstacles and leaves its limbs free to perform tasks. This makes it promising for use in human care, rescue operations and direct interaction with people.
