Learning to walk in arbitrary legged morphologies

In mobile robots, a usually desired feature of the mobility is locomotion, which can be trivial e.g. in the case of wheeled robots, however legged robots often require more elaborate control methods. A common strategy to develop a locomotion controller is to create a model of the robot in simulation, define joint movement parameters and use an optimization method to find a suitable locomotion strategy. Reconfigurable modular robots (RMR) are a special type of robots that have the ability to be assembled into various morphologies according to a set of tasks, among which legged locomotion often is part of. While the controller development described above is also still applicable, it quickly can turn cumbersome as each new morphology requires a new model, parametrization and optimization. Our aim in this work is to explore a learning method coupled together with sensory feedback to develop a generic control strategy (i.e. the “spinal cord”) able to make an arbitrary modular morphology locomote in one shot, i.e. a modular legged structure learns how to move on the fly. Once built, a structure first performs a series of discrete motor actions, so-called “Spontaneous Motor Activities” (SMA), that have been observed to occur during REM-sleep of mammals [1]. We use Hebbian learning as one of the basic biological unsupervised learning method to correlate these discrete motor actions with sensory feedback caused by them [2] to form a rough internal model of the robot with the goal of separating single limbs. We then implement phase oscillators in each limb to synchronize limb movements into an emergent gait through the clever use of force feedback, known as “tegotae”. These processes are deployed into a customized modular robotic platform, and we present validation experiments as well as the first results of the learning procedure.