Towards intuitive rigid-body physics through parameter search

The ability to predict the future location of objects is key for robots operating in unstructured and uncertain scenarios. It is even more important for general purpose humanoid robots that are meant to operate and adapt to multiple scenarios. They need to determine possible outcomes of actions, reason about their effect and plan subsequent movements accordingly to act preemptively. The prediction ability of current robotic systems in is far from that of humans. Neuroscience studies point out that humans have a predictive ability, called intuitive physics, to anticipate the behavior of dynamic environments enabling them to predict and take preemptive actions when necessary, for example to catch a flying ball or grab an object that is about to fall off a table. In this paper, we present a system that learns to predict based on previous observations. First, object's physical parameters are learned through observation using parameter search techniques. Second, the learned dynamic model of objects is used to generate probabilistic predictions through physics simulation. The parameter search update rules proposed, are compared to other approaches from the state-of-the-art in physical parameter learning. Finally, the predictive capability is evaluated through simulated and real experiments.