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

Six-Legged Robot Learns to Walk from a Stick Insect Using AI

Researchers from Tohoku University in Japan and the Vidyasirimedhi Institute of Science and Technology (VISTEC) in Thailand have developed a six-legged robot that learns to walk autonomously by studying stick insect movement patterns1

1

2

2

. The AI-powered system enables the robot to master walking in just one hour and adapt its gait across different surfaces without explicit programming2

2

. This approach could transform how robots navigate disaster zones and unpredictable environments where wheeled systems fail.

Adversarial Inverse Reinforcement Learning Powers Autonomous Walking

The system uses adversarial inverse reinforcement learning (AIRL) combined with proximal policy optimization (PPO) to extract walking strategies from biological coordination patterns1

1

. Researchers trained the AI on movement data from Medauroidea extradentata stick insects, capturing just three or four steps across 18 leg joints2

2

. Instead of copying each motion directly, the system learned underlying coordination principles that govern stable locomotion. During training, one neural network compares the robot's movements with insect demonstrations while another learns a reward structure encouraging behavior similar to the biological example1

1

. Associate Professor Dai Owaki from Tohoku University explains the fundamental shift: "We never told the robot how to walk. We asked what the insect was trying to achieve, and let the robot chase the same thing entirely on its own"2

2

.

Robot Adapts Its Gait Across Uneven Terrain and Recovers from Leg Damage

The six-legged robot demonstrated remarkable adaptability when tested beyond its training conditions. Although trained exclusively on flat-ground walking data, it maintained stable movement across uneven terrain with only slight reductions in forward speed1

1

. The system autonomously adjusted leg timing and coordination, producing a wave-like walking pattern suited to challenging surfaces. When researchers disabled one leg to simulate damage, the controller reorganized how the remaining legs worked together and redistributed the load to maintain stability—responding to a physical change absent from original training data1

1

. The robot learns to walk autonomously three times faster than standard reward-based methods, completing training in 70,000 steps compared to 200,000 steps for conventional velocity-based reward shaping1

1

.

Transferable Learning Accelerates Adaptable Robotics Development

A critical breakthrough involves transferring learned principles between different robot models. The system splits learning into two categories: one holding universal principles applicable to any body, and another containing machine-specific information2

2

. When researchers attempted direct policy transfer to a robot with different body proportions and joint configurations, it failed. However, transferring the AIRL-derived reward structure allowed the new robot to develop coordinated movement independently1

1

. This eliminates the slow, robot-specific design process that requires manual gait programming for each new machine. Owaki notes: "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"2

2

. Preliminary tests on a physical RedMirror robot showed walking and body-coordination patterns qualitatively similar to simulation results1

1

.

Implications for Disaster Response and Future Development

This research addresses a fundamental question: what is walking for, rather than simply how to make robots walk. At disaster zones where conventional wheeled robots cannot navigate debris and unstable surfaces, these adaptive six-legged systems could save lives2

2

. The technology overcomes limitations of conventional robot gait systems that rely on fixed patterns, tuned parameters, and predefined rewards struggling on unfamiliar terrain1

1

. Researchers plan to add memory capabilities so robots can learn and build experience over time, moving toward highly mobile systems for disaster response efforts2

2

. The approach could extend beyond stick insects to other animals with high dexterity, potentially creating a new generation of robots that learn from nature's proven solutions. Watch for developments in transferable learning frameworks that could dramatically reduce development costs and accelerate deployment of adaptable robotics across industries requiring navigation through unpredictable environments.