Online learning of sensorimotor interactions using a neural network with time-delayed inputs

Abstract

The work described here explores an approach for learning online the sensorimotor interaction that a robot has with the world, and the higher-level concepts grounded in this interaction. A type of spatiotemporal connectionist neural network was implemented. In consists of a set of time-delayed input layers which receive both low-level sensor inputs and high-level labels and hypotheses. Each input value activates a range of neurons, based on a Gaussian distribution. A Hebb-like learning rule is used online to associate activations from inputs in the past with activations from inputs in the present. Prediction of future activation is then performed by shifting all inputs one time-step back in time and propagating activation to the present time layers. A simple benchmarking based on a number 8 shape movement with a simulated iCub robot showed good robustness to noise and ambiguity in the trajectories. First results from trials interacting with simulated objects in an imitation learning scenario are also presented. The system was able to learn online and ground labels and hypotheses in the trajectories, although the strength of the predictions was reduced.