LEARNING NAIVE PHYSICS BY VISUAL OBSERVATION: USING QUALITATIVE SPATIAL REPRESENTATIONS AND PROBABILISTIC REASONING

Abstract

Autonomous robots are unsuccessful at operating in complex, unconstrained environments. They lack the ability to learn about the physical behavior of different objects through the use of vision. We combine Bayesian networks and qualitative spatial representation to learn general physical behavior by visual observation. We input training scenarios that allow the system to observe and learn normal physical behavior. The position and velocity of the visible objects are represented as qualitative 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 behavior. We use test scenarios to determine how well the approach discriminates between normal and abnormal physical behavior and actively predicts future behavior. 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.