Dr. Eugene Vinitsky, Apple, University of California (UC) Berkeley
Add Headings and they will appear in your table of contents.
Talk Date and Time: December 8, 2022 at 04:00 pm - 04:45 pm EST followed by 10 minutes of Q&A on Zoom and IRB-5105
Topic: Nocturne: a scalable driving benchmark for bringing multi-agent learning one step closer to the real world
Abstract:
This presentation provides an overview and motivation behind the development of Nocturne, a new 2D driving simulator for investigating multi-agent coordination under partial observability. The focus of Nocturne is to enable research into inference and theory of mind in real-world multi-agent settings without the computational overhead of computer vision and feature extraction from images. Agents in this simulator only observe an obstructed view of the scene, mimicking human visual sensing constraints. Unlike existing benchmarks that are bottlenecked by rendering human-like observations directly using a camera input, Nocturne uses efficient intersection methods to compute a vectorized set of visible features in a C++ back-end, allowing an agent to run at 2000+ steps-per-second. Using open-source trajectory and map data from the Waymo Motion dataset, we construct a simulator to load and replay arbitrary trajectories and scenes from real-world driving data. Using this environment, we benchmark reinforcement-learning and imitation-learning agents and demonstrate that the agents are quite far from human-level coordination ability and deviate significantly from the expert trajectories.
Bio:
Eugene Vinitsky is an incoming Professor at NYU Tandon in Civil and Environment Engineering and a current research scientist at Apple in their special projects group. He received his PhD in controls and optimization at UC Berkeley in Mechanical Engineering. Prior to that, he received his MS in physics from UC Santa Barbara and a BS in physics from Caltech. At UC Berkeley, he focused on scaling multi-agent reinforcement learning to tackle the challenges associated with transportation system optimization. As a member of the CIRCLES consortium, he was responsible for the reinforcement learning algorithms and simulators used to train and deploy energy-smoothing cruise controllers onto Tennessee highways. In the past he has spent time at Tesla Autopilot, DeepMind and was a visiting researcher at Facebook AI. His research has been published at ML venues such as CORL, neurIPS, and ICRA and at transportation venues such as ITSC. He is the recipient of an NSF Graduate Student Research Award, a two time recipient of the Dwight David Eisenhower Transportation Fellowship, and received an ITS Outstanding Graduate Student award.