RoboticsRobot chooses from a menu of walking styles to escape trouble
When a newly developed robot finds it cannot move freely, it scans through the many walking gaits it has taught itself and selects the best for the terrain; this means it can free itself should it get stuck; the robot has six triple-jointed legs each with several sensors; the sensors feed information to the neural network, which then determines the most appropriate gait for the terrain, and adjusts the robot’s eighteen motors accordingly
A six-legged robot learns different walking styles, which it can then use to adapt to tricky terrain or even flee from the first signs of trouble. A six-legged robot with a “panic mode” is proving to be a whiz at locomotion. When it finds it cannot move freely, the panicky droid scans randomly through the many walking gaits it has taught itself and selects the best for the terrain. That means it can free itself should it get stuck.
Paul Marks writes that getting robots to choose the right gait on differently textured surfaces and at varying inclinations is tough. Some robots use preprogrammed gaits, while others use software routines called genetic algorithms (GAs) to evolve the best gait on the fly. Both these methods need a lot of onboard computer power.
Silke Steingrube of the Bernstein Center for Computational Neuroscience in Göttingen, Germany, and colleagues took a different tack. They opted instead for a simple computer called a neural network, a computer system that uses feedback on previous decisions to learn from its experiences. Their robot has six triple-jointed legs each with several sensors. These feed information to the neural network, which then determines the most appropriate gait for the terrain, and adjusts the robot’s eighteen motors accordingly.
Using its raft of sensors — which detect foot-contact pressure, light, sound, heat and the robot’s inclination — the robot can select the correct gait for uphill, downhill, and various types of rough ground. By programming the robot to adopt the most energy-efficient gait possible, the researchers ensured it would switch gaits whenever its incline sensors were triggered. In tests, the robot taught itself 11 different walking styles. “The technique should work equally well in four-legged, six-legged or even wheeled robots,” says Steingrube.
Marks writes that it has a flight reflex, too: if a rear sensor detects, say a very high temperature, it interprets this as a threat. “The neural network generates a fast, wave-like gait that is appropriate for running away,” says Steingrube.
If the robot gets into difficulty, with a foot stuck in a hole, say, a number of sensors are stimulated. This creates a large input signal, which induces an unpredictable, chaotic output from the neural network, causing it to randomly choose one of its eleven gaits. In other words, the robot cycles through its repertoire until it frees itself.
Chris Melhuish, >head of the Bristol Robotics Lab in the United Kingdom, is impressed. “If you get stuck, going into what’s effectively a ‘twitching’ mode like this could indeed be useful,” he says. “It would be good to see if they can adapt this to help robots that have been damaged – lost a leg perhaps, or with a motor that is underperforming.”
At the Robert Gordon University in Aberdeen, UK, roboticist Chris Macleod has been using GAs to evolve robot gaits. “This chaotic mechanism is an interesting idea and certainly merits more experimentation because many biological neural networks, like those in the autonomic nervous system, are known to exhibit chaotic or semi-chaotic behavior,” he says.
