Our aim is to better understand the action selection process of intelligent systems by looking at their ability of internal prediction. In robotic systems, one problem is to generate meaningful robot behaviour with a very small and simple set of trained motions. An additional problem is to compensate for incomplete sensory data while generating behaviour. We propose a new predictive action selector to contribute to the solution of these problems. Our action selector predicts task-relevant feature and motion sequences, and uses the prediction results to select the robot action. We validate our implemented model on a humanoid robot. The robot generates meaningful behaviour composed out of very simple and few trained motions, and at the same time it compensates for incomplete sensory data such as temporary loss of task-relevant visual features.
Abstract-Developmental robotics suggests that the forward and inverse kinematics should be learned through a sensorymotor mapping, instead of being programmed in advance. Motor babbling and goal babbling are two common approaches to generate training samples used to acquire a sensory-motor mapping. Motor babbling typically needs a considerable amount of training data and time to acquire a sufficient mapping, while goal babbling poses difficulties on how to select appropriate goals. In this paper, we propose a neurobiologically-inspired system to progressively learn a sensory-motor mapping bootstrapped from a simple constrained DOF exploration, which generates much less training data than motor babbling. Our proposed system is designed according to two neurobiologicallyinspired paradigms: spatiotemporal prediction and uniformity. The spatiotemporal prediction capability facilitates the acquisition of sensory-motor mappings with less amount of training data on the one hand, and facilitates robust behaviour on the other hand. The uniform system design structure is the foundation for building a scalable architecture for cognitive development. We use an improved version of our predictive action selector (PAS) as building block of our system. We validate a PAS on a 2 DOF robot head where the robot learns object tracking and evading. Then we validate a second PAS on a 5 DOF arm where it learns reaching.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.