Using deep reinforcement learning (DRL), the EPFL robot has successfully learned to transition from trotting to pronking, a leaping gait used by animals like springbok and gazelles, in order to navigate a difficult terrain with gaps ranging from 14-30cm. Conducted by the BioRobotics Laboratory in EPFL’s School of Engineering, the study provides new insights into the reasons behind such gait transitions in animals.
Previous research has suggested that energy efficiency and musculoskeletal injury avoidance are the main reasons for gait transitions. However, recent studies have proposed that stability on flat terrain may be more crucial. The team, led by PhD student Milad Shafiee, aimed to explore a new hypothesis: viability or fall avoidance. Through DRL, they trained a quadruped robot to navigate different terrains and discovered that gait transitions occurred to maintain viability, with energy efficiency not necessarily being improved.
The team revealed that viability was the driving force behind gait transitions, with energy efficiency being a consequence rather than a primary factor. When animals navigate challenging terrain, their priority is likely to avoid falling, followed by energy efficiency.
To model locomotion control in their robot, the researchers considered the brain, spinal cord, and sensory feedback from the body. By using DRL to train a neural network to imitate the brain signals transmitted by the spinal cord, the team found that viability was the only goal that led to automatic gait changes in the robot.
This study represents the first learning-based locomotion framework where gait transitions emerge spontaneously during the learning process. The researchers aim to continue their work by conducting experiments in various challenging environments to further understand animal locomotion and potentially reduce reliance on animal models for biological research.
