Reminder about this talk today.

---------- Forwarded message --------- From: Mikhail Bessmeltsev bmpix@iro.umontreal.ca Date: Thu, Apr 28, 2022 at 9:16 AM Subject: [DIRO Colloque] Rappel: Avr 28, Avi Singh, Augmenting Robotic Reinforcement Learning with Prior Datasets To: seminaires@iro.umontreal.ca

Aujourd’hui! On Apr 21, 2022, 10:03 PM -0400, Mikhail Bessmeltsev bmpix@iro.umontreal.ca, wrote:

Bonjour à tous,

Notre prochain colloque aura lieu dans une semaine ! Tout le monde est bienvenue, comme toujours. La présentation sera en anglais.

*Quand*: 28 avril 2022, 15h30-16h30 HNE (EST) *Où*: https://umontreal.zoom.us/j/82974092228?pwd=UEVFc0VTVGNMZk9FdVNKRUdrSGNXZz09 Zoom Meeting ID: 829 7409 2228, Passcode: 409200 *Qui*: Avi Singh, Google Brain *Titre*: Augmenting Robotic Reinforcement Learning with Prior Datasets *Résumé*: Reinforcement learning provides a general framework for flexible decision making and control, but requires extensive data collection for each new task that an agent needs to learn. Further, policies learned in such fashion are often too brittle, and do not generalize to new scenarios. In this talk, I will present a few ways in which robotic reinforcement learning can be improved with help from previously collected diverse, multi-task interaction datasets. First, I will present a method for pre-training RL agents using data from a wide range of previously seen tasks, and show how this pre-training can accelerate learning of new tasks. Then, I will present how prior datasets can be utilized to help achieve better generalization to novel scenarios.

*Bio*: Avi Singh is a Research Scientist at Google Brain, where he works at the intersection of machine learning and robotics. In particular, his research focuses on making reinforcement learning techniques amenable to real world robotics. He received a PhD in Computer Science from UC Berkeley in 2021, where he was advised by Sergey Levine. His dissertation focused on learning reward functions from small human-provided datasets, and the role of prior data in robotic reinforcement learning. He has spent time at Google X, Cornell University and Virginia Tech, and received his undergraduate degree from the IIT-Kanpur.

Cordialement, Mikhail Bessmeltsev