Towards Learning Naive Physics by Visual Observation: Qualitative Spatial Representations

Abstract

Autonomous robots are unsuccessful at operating in complex, unconstrained environments. They lack the ability to learn about the physical behaviour of different objects. We examine the viability of using qualitative spatial representations to learn general physical behaviour by visual observation. We combine Bayesian networks with the spatial representations to test them. We input training scenarios that allow the system to observe and learn normal physical behaviour. The position and velocity of the visible objects are represented as discrete states. Transitions between these states over time are entered as evidence into a Bayesian network. The network provides probabilities of future transitions to produce predictions of future physical behaviour. We use test scenarios to determine how well the approach discriminates between normal and abnormal physical behaviour and actively predicts future behaviour. We examine the ability of the system to learn three naive physical concepts, ‘no action at a distance’, ‘solidity’ and ‘movement on continuous paths’. We conclude that the combination of qualitative spatial representations and Bayesian network techniques is capable of learning these three rules of naive physics.