I am interested in endowing embodied agents with the ability to reason about the world around them,through physical scene understanding and continually improve their model of theworld through lifelong learning.
We conduct over 700 experiments to analyze the effects of using various pre-trianing datasets, and self supervised reinforcment learning pipelines, on the final performance of computer vision encoders on a variety of end tasks.
We introduce a new library and framework for conducing reinforcment learning experiments in PyTorch. By providing many built-in training algorithms and utilities and allowing for easy expansion, we hope to reduce the developement overhead and improve consistency of results across projects.
