A neural network method of learning human motion by observation in operational space

Learning by observation is a useful goal for humanoid robots that strive for human-like motion. Over time the movements and actions of these robots are likely to require modification for different scenarios and environments. Therefore being able to ‘teach’ the robot new tasks and motions without tedious programming is highly useful. The operational space formulation is a highly elegant robot manipulator control method that has parallels with human motion strategies, with prioritization of task achievement and redundant degrees of freedom control via effort optimisation. In this paper, we combine a novel method of learning by observation with the operational space framework. This has created a learning controller that is able to generalise a small set of example reaching motions in order to reach to new targets while maintaining human like motion rajectories in the posture space via a novel sliding mode optimal controller. The trajectories are executed on a (dynamically) simulated robot arm which performs human like reaching to arbitrary locations. Trajectories are minimally encoded as the coefficients of fitted polynomials and the learning is performed by three neural networks that learn and generalise. Initial results are presented for this early stage of research on this proposed framework.