Structured Policy Representation. Stability in arbitrarily conditioned dynamic systems

The poster presents a new family of deep neural network-based dynamic systems and how these dynamics can be used as structured robot policies.  Additionally, the poster details the concept of global stability, as one of the most important and straightforward inductive biases which allows to impose reasonable behaviours outside the region of the demonstrations.  

In the work, a novel set of Conditional Invertible Neural Networks were developed. Invertible neural networks are a special type of networks that allow to compute the inverse and thus, we exploited this property to learn globally stable conditioned dynamics. The developed networks were applied for Imitation Learning problems. In our work, we presented a toy-task in which a limit cycle vector field was adapted given a phase variable, a 2D obstacle avoidance task, in which the robot should arrive to a target while avoiding a set of possible obstacles and finally, a robot learning task, in which the robot should learn how to pour in a randomly position pot from data.

The obtained results would be exploited in Sharework EU project. A software module on primitives learning is dedicated to developing new movement primitives that will allow to model both human and robot movements for human motion prediction or robot motion control.