An EPFL robot reveals why animals change their appearance

An EPFL robot reveals why animals change their appearance
An EPFL robot reveals why animals change their appearance

The robot started jumping like gazelles to avoid the holes.

EPFL/BioRob-CC-BY-SA 4.0

Trained using machine learning by EPFL scientists, a quadruped robot spontaneously changed its gait in an attempt to avoid falls. This is an important step for robotics specialists and biologists interested in animal locomotion.

Using a form of machine learning called deep reinforcement learning (DRL), the EPFL robot notably learned to switch from trotting to stotting (a behavior in which animals like springboks and gazelles leap with an arched back) to move on difficult terrain with holes. Led by the Biorobotics Laboratory of the Faculty of Engineering Sciences and Technology at EPFL, this study provides a better understanding of why and how such changes in gait occur in animals.

“According to previous research, animals change their gait to conserve energy and avoid musculoskeletal injuries. More recently, biologists have argued that stability on flat ground may be a more important factor. But experiments with animals and robots have shown that these assumptions are not always valid, particularly on uneven ground,” reports Milad Shafiee, doctoral student and lead author of an article published in the journal Nature Communications.

Milad Shafiee and co-authors Guillaume Bellegarda and Auke Ijspeert, head of the Biorobotics Laboratory, were therefore interested in a new parameter that could explain these changes in pace: viability, or the prevention of falls. To test this hypothesis, they trained a quadruped robot to traverse various terrains using deep reinforcement learning.

Jumps to avoid holes

On flat ground, they found that each gait had different levels of robustness to random pushes and that the robot switched from walking to trotting to maintain viability, as quadrupedal animals do when they accelerate. And when confronted with successive holes of 14 to 30 cm in the experimental surface, the robot spontaneously switched from trotting to stotting to avoid falls. Furthermore, viability is the only factor that has been improved by these changes in pace.

“We showed that on flat ground and on difficult terrain, viability leads to changes in pace, but that energy saving is not necessarily better,” explains Milad Shafiee. It therefore seems that energy saving, which was previously considered a factor explaining these changes, could be more of a consequence. When an animal is moving through difficult terrain, it is likely that its priority is not to fall. Saving your energy would come next.”

The robot automatically changes its pace

To model the control of their robot’s movements, the scientists took into account the three interacting elements that control the animals’ movements: the brain, the spinal cord and sensory feedback from the body. They used deep reinforcement learning to train a neural network to mimic the transmission of brain signals from the spinal cord to the body as the robot navigated an experimental terrain. Next, the team assigned different weights to three possible learning goals: energy saving, force reduction, and viability. A series of computer simulations revealed that, of these three objectives, viability was the only one that caused the robot to automatically change its pace, without instruction from scientists.

The team emphasizes that these observations represent the first learning-based locomotion framework in which gait changes appear spontaneously during the learning process, as well as the most dynamic crossing of consecutive gaps this large for a robot quadruped. “Our bio-inspired learning architecture demonstrated the agility of a state-of-the-art quadruped robot in difficult terrain,” says Milad Shafiee.

The scientists want to expand on their work by conducting other experiments that place different types of robots in a wider variety of harsh environments. They hope that ultimately, their work will not only make it possible to understand animal locomotion, but also to generalize the use of robots for biological research, by reducing the use of animal models and the ethical problems associated with them.




PREV LIVE – War in Gaza: an Israeli operation in Rafah would be an “unspeakable tragedy”, worries the UN
NEXT Two Rouge et Or players drafted by the Montreal Alouettes