Physical Society Colloquium

Physically Intelligent Robotics

David Meger

School of Computer Science
McGill University

There is a great need for robots that can operate in dynamic environments, such as oceans and rivers, but the physics involved are often difficult to model (for a computer scientist) and rapidly changing. This calls for physical intelligence: the ability of a system to autonomously learn about present physical conditions, behave productively according to this knowledge and adapt as circumstances change. I will describe research at McGill that has developed a simple physical intelligence for a swimming robot with six flippers. This method begins without any given model of the robot's hydrodynamics and utilises observed experience data to learn predictive models. In order to actively gather additional data and to improve behaviour, both the current model's prediction and its uncertainty are used to optimise the parameters of a neural network control policy. This approach achieves similar performance to the best human-crafted controllers in around 8 trials of simple swimming behaviours such as a corkscrew pattern.

The second half of the talk will discuss a variety of approaches to extend the “intelligence” of these basic techniques. Rather than re-learning models and controllers when small changes occur, we have developed methods to rapidly adjust to new circumstances by automated analysis of the differences in observed dynamics. This approach also allows the expensive computations of initial learning to occur in a crudely approximate software simulation, leaving only rapid refinement steps to be carried out on hardware robots. I will finally describe recent progress that allows robots to imitate human behaviours, on dynamics models that learn even faster by using physically-motivated basis functions, and I will show videos of robots swimming in warm places like the Caribbean Ocean to combat the rainy Montreal weather.