Title: Emergence of Intelligent Behavior in Embodied AI Agents from Massive Scale Simulation
Date: Friday, October 15, 2021
Time: 11:30am-1:00pm ET
Location (virtual): https://bluejeans.com/143359383/4938

Erik Wijmans
PhD Student in Computer Science
College of Computing
Georgia Institute of Technology

Committee
Dr. Irfan Essa (Advisor, School of Interactive Computing, Georgia Institute of Technology)
Dr. Dhruv Batra (Co-advisor, School of Interactive Computing, Georgia Institute of Technology)
Dr. Sonia Chernova (School of Interactive Computing, Georgia Institute of Technology)
Dr. Vladlen Koltun (Apple)
Dr. Gregory Wayne (DeepMind)
Dr. Vincent Vanhoucke (Google)

Abstract
A long term goal in artificial intelligence (AI) research is to build intelligent agents, deployed on robots, that are capable of performing goal-oriented tasks that require interaction with the environment, e.g. "bring my my laptop". Recent years have seen progress towards this goal via the paradigm of training virtual robots, embodied agents, in simulation. 

In this thesis we will argue that intelligence emerges from massive scale training in simulation. We present 1) Fast and efficient simulators. Training at massive scale a requires fast and resource efficient simulator. We present the habitat simulator. 2) Scalable reinforcement learning systems. We present Decentralized Distributed PPO (DD-PPO) that scales learning to multiple GPUs and machines. In proposed work, we propose to build a hybrid synchronous and asynchronous system that improves scaling and throughput for simulators with highly variable simulation time, as is common with physics simulators. 3) Analysis of emergent intelligent behaviors. Finally we demonstrate that intelligent behaviors emerge with massive scale simulation training. We find that agents are capable of near-perfect navigation in unseen environments and emergently learn to build maps. When combined with decades of research in animal navigation, this suggests that map building may be a natural solution to navigation. In proposed work, we will analyze agents that are trained to navigate to a location and pick an object up.