A team of researchers have designed a four-legged robot that is capable of spontaneously transitioning between different modes of locomotion. Thanks to a form of machine learning known as deep reinforcement learning (DRL), this innovative robot can switch between walking, trotting, and pronking — an arched-backed leap common in animals such as gazelles and springboks. The flexibility of the transitions allows the robot to navigate complex terrains smoothly and further provides insights into how animals alter their gait to adapt to diverse environments.
The study conducted by the leading group at Ecole Polytechnique Fédérale de Lausanne's (EPFL) BioRobotics Laboratory illustrates how viability, i.e., the potential for fall avoidance, motivates gait transition in a manner reminiscent of quadruped animals. In contrast to traditional hypotheses proposing energy efficiency and injury avoidance as main reasons for gait transitions, the research team posits that stability in flat terrain could be an even more pressing concern.
To fully investigate this hypothesis, the team used DRL to train their four-legged robot to traverse a multitude of terrains. In simulating scenarios with different levels of pushing forces, they observed the robot transitioning from walking to trotting, displaying the inherent robustness of biomechanical movement. Tackling terrains with successive gaps presented a further challenge, yet the robot smoothly shifted from trotting to pronking to prevent falling. The team found that viability was indeed the primary factor enhanced by these gait transitions.
EPFL Ph.D. student and first author on this study, Milad Shafiee, states, “Our observations emphasise that while moving across challenging terrains, avoiding a fall is likely an animal’s first priority, followed by energy efficiency.” In contrast to previous beliefs, energy efficiency might simply be a subsequent benefit rather than a primary stimulus for gait transition.
The team modelled the robot’s locomotion control based on the three interactive elements driving animal movement — the brain, the spinal cord, and sensory feedback from the body. They trained their robot’s neural network to channel brain signals to the body, mimicking the spinal cord’s functions. The team factored in energy efficiency, force reduction, and viability when deciding on the goals for learning. Their computer simulations showed that viability alone prompted gait changes in the robot.
Shafiee also highlights that the team’s findings mark the first learning-based locomotion model where gait transitions occur spontaneously during the learning process. It also represents the most dynamic way a quadruped robot has been documented to tackle substantial and consecutively placed gaps.
The team's research efforts and resultant bio-inspired learning architecture have paved the way for further experimental exploration and showcase cutting-edge quadruped robot agility on varied terrains and challenges. Their aim is to conduct further studies involving different types of robots and environs, expanding our understanding of locomotion and reducing animal model dependencies, thus addressing ethical concerns.
Disclaimer: The above article was written with the assistance of AI. The original sources can be found on ScienceDaily.
