Six-Legged Robot Masters Walking by Learning from Stick Insect in Just One Hour

Stick insect-inspired robot adapts its gait across difficult terrain

Researchers have developed a six-legged robot that can learn to walk by mimicking a stick insect's movement patterns. The AI-powered system enables the robot to adapt its walking strategy across different surfaces and challenging environments. Researchers from Tohoku University in Japan and VISTEC in Thailand say the approach could help robots move more effectively through unpredictable terrain. The technology could eventually support robotic systems designed to operate in disaster zones and other environments where conventional wheeled robots may struggle to navigate. Researchers have developed a new AI-based method that allows a six-legged robot to learn walking strategies from a stick insect. Instead of manually programming how each leg should move, the system studies biological movement data and uses it to learn the underlying coordination patterns needed for stable locomotion. The approach uses adversarial inverse reinforcement learning (AIRL) to learn walking strategies from stick insects. Researchers trained the system on flat-ground walking data from Medauroidea extradentata, capturing movements across 18 leg joints. Instead of copying each motion, the AI learned the underlying coordination principles. This could overcome the limits of conventional robot gait systems, which rely on fixed patterns, tuned parameters, and predefined rewards that may struggle on unfamiliar terrain. The new system combines AIRL with proximal policy optimization (PPO), a reinforcement-learning method. During training, one neural network compares the robot's movements with the insect demonstrations, while another learns a reward structure that encourages behavior similar to the biological example. According to a statement by researchers, the robot's controller uses information such as body orientation, joint angles, and whether each leg is touching the ground. It then produces commands for the robot's 18 joints. This allows the system to learn coordinated leg movements without researchers having to explicitly program a particular gait. The researchers found that the learned controller could produce stable walking that resembled the coordination seen in stick insects. More importantly, the system was able to adapt when conditions changed. Although it was trained using only flat-ground walking data, the robot continued to walk on uneven terrain. Its body experienced greater movement as the ground changed, but it remained stable, and its forward speed fell only slightly. The system also changed the timing and coordination of its legs, producing a more wave-like walking pattern suited to uneven surfaces. The robot also demonstrated an ability to cope with the loss of a leg. Researchers disabled one of its six legs to simulate damage. Instead of simply continuing with the original walking pattern, the controller reorganized how the remaining legs worked together and redistributed the load to maintain stability. This showed that the AI could respond to a physical change that was not present in its original training data. Another key result was the ability to transfer the learned reward structure to a different robot model with different physical properties. Directly transferring the original walking policy failed because the second robot had different body proportions, joint configurations, and motor characteristics. However, transferring the AIRL-derived reward allowed a new policy to learn coordinated movement on the different robot. The combined AIRL and forward-velocity rewards reached a useful walking strategy in 70,000 training steps, compared with 200,000 steps when using velocity-based reward shaping alone. The researchers also carried out a preliminary test on a physical RedMirror robot. The robot showed walking and body-coordination patterns that were qualitatively similar to those seen in simulation. The results suggest that biological movement data could provide a practical way to build more adaptable robot locomotion systems, particularly for machines expected to operate across changing terrain or after mechanical damage.

Six-Legged Robot Learns to Walk from a Stick Insect | Newswise

Newswise -- Do you ever see an ant skittering across your kitchen counter and wonder how this little bug overcame all the obstacles in your home just to get there? Insects use their tiny nervous systems to walk with surprising dexterity - a skill that could be transferred to robots looking to do more productive things than just finding breadcrumbs in your kitchen. An international team of researchers led by Tohoku University in Japan and the Vidyasirimedhi Institute of Science and Technology (VISTEC) in Thailand trained AI on the walking cycle of a stick insect to find the best strategy for walking on different surfaces. This was then successfully implemented in a six-legged robot that learned to walk in just an hour. In addition, the robot was able to navigate uneven terrain and adapt to a missing limb. At treacherous disaster sites where wheeled robots are unable to enter, these adaptive six-legged robots could be genuine life-savers. Stick insects, named for their uncanny resemblance to a small stick, are a classic model for walking research. The team used an open data source covering just three or four steps of the insect. From that data, an AI worked out two things at the same time. First, what the insect seems to be aiming for when it walks (called the "reward"). Second, how the legs should move to achieve it. Most work asks how to make a robot walk. This study asks what walking is for. "We never told the robot how to walk," explains Dai Owaki, Associate Professor at Tohoku University. "We asked what the insect was trying to achieve, and let the robot chase the same thing entirely on its own." The robot learned to walk three times faster than with a standard reward. Giving the robot specific instructions about how to move each leg is a slow design process that has to be redone for every new robot. In contrast, this approach removes that step by giving the robot an animal teacher to learn from. The robot learns two categories: one part holds what is true for any body. The other part holds information that belongs to one particular machine. This split allows for the result to be moved to other robots. This means each new machine doesn't have to start from zero, which could lead to cheaper and faster production. "It's remarkable that a few steps from a single stick insect were enough to find a principle that works on a machine five times its size," says Owaki. The study investigated an inverse reinforcement learning method where AI isn't explicitly told how to walk - it learns from trying to understand the goal of an ideal example. In the current study, it is a stick insect, but other animals with high dexterity could also serve as models. The researchers note that adding memory so the robot can learn and build experience over time will be their next step towards developing highly mobile robots who may one day aid in disaster response efforts.