Robots Can Now Navigate on Rugged Terrain Without Onboard Cameras

Researchers at Carnegie Mellon University and the University of California, Berkeley recently developed a system for an affordable, small-legged robot, enabling it to travel anywhere. In this case, they can climb up and down stairs that almost match their height, work in darkness, scale rocks and curbs, walk across gaps, and traverse slippery, rough, steep, and varied terrain.

The team tested their system at public parks, where it walked on slippery surfaces, across stepping stones, and climbed stairs. It uses the small onboard computer and vision to navigate quickly across difficult terrain. The robot underwent training with 4,000 copies of itself in a simulator, practicing the ability to walk and climb on rough ground. A neural network installed on the actual robot contains all the trained motor skills.

Rather than relying on mapping and planning techniques seen in traditional robots, this system routes the vision inputs to the robot's control. So the robot moves a certain way based on its vision. This approach means the team doesn't need to instruct how the legs move. As a result, the robot quickly adapts to the terrain on its path and traverses on the surface. Additionally, the system is 25 times more affordable, potentially making robots more available than other solutions since it trains via machine learning.

The researchers found inspiration from humans and animals to provide the robot with better movement. They previously demonstrated that robots operating without cameras could maneuver on difficult terrain, but it improved after introducing vision to the system.

Other aspects of the system were nature-inspired as well. For example, a small robot adopted similar movements humans use to step over high obstacles. By doing so, it can scale stairs and obstacles that almost match its height. Humans use abduction and adduction to move their legs out to the side, enabling them to lift their legs for ledge or hurdle scaling. This robotic system achieves the same hip abduction movement to overcome obstacles that otherwise interfere with advanced-legged robots.

The team was also inspired by cats. While moving through obstacles, its hind legs avoid objects without relying on vision. So the system's memory allows the rear legs to remember what its frontal camera saw, traversing for obstacle avoidance. Overall, this study offers potential solutions toward solving challenges for legged robots, making them more convenient for home use in the future.