Robot Learns Insect Gait in One Hour to Traverse Rough Terrain

When an earthquake turns flat roads into rubble, or when the rugged Martian landscape stops all wheeled vehicles, who will reach the trapped? An international team from Tohoku University in Japan and VISTEC in Thailand is seeking answers from a slender insect.

Their biomimetic hexapod robot, designed after stick insects, offers a new approach to locomotion in challenging environments.

The team chose the stick insect as their model. These insects possess six slender legs, allowing them to move flexibly through branches, leaves, and grounds, with coordination far surpassing most current multi-legged robots. However, entomologists have long struggled to precisely identify the mathematical principles of their leg coordination, making it difficult to translate observed behavior into programmable rules.

Associate Professor Fumihiko Owaki, corresponding author of the study, stated, "We did not tell the robot how to walk from the beginning. We only let it understand what the stick insect wants to achieve, and then let the robot autonomously pursue that goal." The team employed adversarial inverse reinforcement learning, a method that automatically identifies reward patterns for safe footholds and continuously adapts to environmental changes.

Experimental results confirmed the efficiency of this approach. The hexapod robot only needed brief observation of stick insects walking on flat ground to autonomously learn to navigate various terrains, all within less than an hour. More importantly, the learned coordination network showed good transferability, serving robots of different body sizes with up to fivefold differences without cumbersome adjustments.

The underlying logic is not to copy the insect's specific movements, but to distill the fundamental principles of its motion control. Traditional biomimetic robots typically require walking programs tailored to specific body types and terrains; changes in environment or body structure necessitate algorithm rewrites. Owaki's team demonstrates that goal-based inverse reinforcement learning can generate more universal motion control frameworks.

The researchers point out that this technology could play a crucial role in disaster scenarios. Rubble environments lack flat surfaces, and rescue robots may face sudden issues like broken limbs, slippery surfaces, or temporary obstacles. If a learning system understands the fundamental goal of walking, rather than memorizing fixed gaits, it can potentially adjust its posture and continue moving even when structurally damaged.

Japan, being earthquake-prone, has an urgent need for disaster robots. This study at Tohoku University has received funding from the Japan Society for the Promotion of Science (JSPS). The team is also exploring extending this learning framework to other hexapods and even quadrupeds.