Overview

As reinforcement learning agents are tasked with solving increasingly challenging tasks, the ability to inject prior knowledge about the structure of the world and about effective behaviors is becoming increasingly important. In this work we consider how information and architectural constraints can be combined with ideas from probabilistic modeling to learn behavior priors that capture the common movement and interaction patterns that are shared across a set of related tasks or contexts.

We extend these ideas to latent variable models and consider a formulation to learn hierarchical priors that capture different aspects of the behavior in reusable modules. In what follows we include a video summary of some of the intuitions behind our approach. We also include videos of all tasks considered in the main text for convenience.

Summary

Task Videos

Locomotion Tasks

Sequential Navigation Tasks