Locomotion in Modular Robots based on Central Pattern Generators

We address the problem of learning to locomote in modular robotic systems, i.e. systems made from multiple homogenous or heterogenous and highly autonomous modules being assembled into a bigger robotic structure, other than monolithic robots (such as humanoid robot systems or predefined quadruped robots). We are particulary interested in online learning of robust and adaptive locomotion for arbitrary robotic structures. We are using Central Pattern Generators (an approach inspired by nature coping with redundancies in animal bodies and the easy generation of locomotion patterns) in combination with a gradient-free optimization method (Powell’s method). We applied our approach on three different robot configurations, interesting locomotion modes are obtained after running the optimization for less than 60min. Our CPG model can be implemented easily in a distributed system among the modules with a low demand for computational power, it exhibits limit cycle behaviour (temporary perturbations are rapidly forgotten), the CPG produces smooth trajectories when control parameters are abruptly changed, and it is robust against imperfect communication between modules. As a testing and application platform we have developed the modular robotic system YaMoR. Among other key features it provides wireless communication among the several YaMoR units based on a specially developed Scatternet Protocol (SNP). As for the future use as an autonomous reconfigurable system (SelfReconfigurable Modular Robots) we are exploring reconfiguration strategies based on a central approach using graph signatures. Introduction As for all mobile robots, one of the key features for a modular robot is robust locomotion. Designing efficient locomotion controllers for such modular systems is however a difficult and unsolved problem. Their locomotion control suffers from all the traditional difficulties of locomotion control in robots with multiple degrees of freedom. Additional difficulties arise due to their modular structure. (Selfreconfiguring) modular robots are not necessarily optimized for a specific locomotion tasks (sub-optimal in terms of balance and mass distribution, sufficient actuation, alignment and amount of degree of freedom). Therefore the control of locomotion requires multi-dimensional coordinated rhythmic patterns that need to be correctly tuned such as to satisfy ∗To whom correspondence should be addressed. θ i