Emergence of Intelligent Navigation Behavior in Embodied Agents from Massive-Scale Simulation

Thumbnail Image
Wijmans, Erik
Batra, Dhruv
Essa, Irfan
Associated Organization(s)
Organizational Unit
Supplementary to
The goal of Artificial Intelligence is to build ‘thinking machines’ that ‘use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.’ In this dissertation, we will argue that the intelligence required for this goal emerges from massive-scale simulation. We will show a specific case: that intel- ligent navigation behavior emerges from massive-scale simulation and deep reinforcement learning. Towards this end, we introduce Decentralized Distributed PPO (DD-PPO), a method that scales reinforcement learning to multiple GPUs and machines. We use DD-PPO to train agents for PointGoal navigation (e.g. ‘Go 5 meters north and 10 meters east relative to start’) for the equivalent of 80 years of human experience. This massive-scale training results in near-perfect autonomous navigation in an unseen environment without access to a map. We then examine the inner workings of special case of PointGoalNav agents. We find that (1) their memory enables shortcuts, i.e. efficiently travel through previously unexplored parts of the environment; (2) there is emergence of maps in their memory, i.e. a detailed occupancy grid of the environment can be decoded from it. We then introduce Variable Experience Rollout (VER), a method that efficiently scales reinforcement learning on a single GPU or machine. We use VER to train chained skills for mobile manipulation. We find a surprising emergence of navigation in skills that do not ostensibly require any navigation. Specifically, the pick skill involves a robot picking an object from a table. During training, the robot was always spawned close to the table and never needs to navigate. However, we find that if navigation actions are part of the action space, the robot learns to navigate then pick an object in new environments with 50% success, demonstrating surprisingly high out-of-distribution generalization.
Date Issued
Resource Type
Resource Subtype
Rights Statement
Rights URI