Guillaume Bellegarda and Auke Ijspeert have used deep reinforcement learning and networks of oscillators that model spinal pattern generators to teach a four-legged robot 11 gaits across difficult terrain, according to a paper on arXiv. Each gait comes with its own style, such as body height and foot clearance.

A second policy chooses among the gaits to meet a goal set by the user, such as low energy use, stability or smooth joints. Each gait proved best for at least one combination of terrain, goal and speed. Without further training, the method coped with new gaits and styles, terrain far outside what it was trained on and quadrupeds five times heavier than the training robot.

Source: arXiv Robotics (cs.RO)

Background

Guillaume Bellegarda

Robotics researcher

Robotics researcher.