Researchers at EPFL’s BioRobotics Laboratory used deep reinforcement learning (DRL) to train a quadruped robot to transition between different gaits to navigate challenging terrain with gaps ranging from 14-30cm. The study aimed to explore the reasons why gait transitions occur in animals and how these transitions can be applied to robotics. Previous research has suggested that gait transitions are driven by factors such as energy efficiency, musculoskeletal injury avoidance, and stability on flat terrain. However, the researchers proposed a new hypothesis that gait transitions are primarily driven by viability, or fall avoidance.
The team found that on flat terrain, different gaits showed varying levels of robustness against random pushes, and the robot switched from a walk to a trot to maintain viability, similar to quadruped animals accelerating. When faced with consecutive gaps, the robot spontaneously transitioned from trotting to pronking to avoid falls, demonstrating that viability was the main driver of gait transitions. This challenges previous beliefs that energy efficiency was the primary driver of gait transitions, suggesting that animals prioritize avoiding falls over energy conservation when navigating challenging terrain.
To model locomotion control in their robot, the researchers considered the brain, spinal cord, and sensory feedback as the three elements that drive animal movement. They used DRL to train a neural network to imitate the transmission of brain signals to the body via the spinal cord. By assigning different weights to learning goals such as energy efficiency, force reduction, and viability, the researchers found that viability was the only goal that prompted the robot to spontaneously change its gait without external instruction. This bio-inspired learning architecture led to dynamic gait transitions in the robot during the learning process.
The team highlighted that their work represents the first learning-based locomotion framework where gait transitions emerge spontaneously, as well as enabling a quadrupedal robot to cross challenging terrain with large consecutive gaps. They hope to expand on their research by conducting further experiments with different types of robots in a wider variety of challenging environments. By elucidating animal locomotion and applying these principles to robotics, the researchers aim to reduce reliance on animal models for research purposes, addressing ethical concerns associated with animal testing.
Overall, the study sheds light on the importance of viability in driving gait transitions in animals and robots navigating challenging terrain. By using deep reinforcement learning and a bio-inspired learning architecture, the researchers were able to train a quadruped robot to transition between gaits in response to different terrains and obstacles. This work paves the way for more agile and adaptive robotic systems that can navigate complex environments and contribute to advancements in both robotics and biological research.
