NAVER LABS Europe seminars are open to the public. This seminar is virtual and requires registration.
Date: 22nd February 2021, 09:00 am (GMT +01.00)
Scaling reinforcement learning to large interactive environments and long-horizon robotic tasks
Speaker: Fei Xia is a fifth-year PhD student at Stanford University. He is advised by Silvio Savarese and Leo Guibas. During his PhD, he interned with Dieter Fox at Nvidia, and Alexander Toshev and Brian Ichter at Google. He obtained his bachelor’s degree from Tsinghua University in 2016. His mission is to build intelligent embodied agents that can interact with complex and unstructured real-world environments, with applications to home robotics. He approaches this problem from 3 aspects: 1) Large scale and transferrable simulation for Robotics. 2) Learning algorithms for long-horizon tasks. 3) Combining geometric and semantic representation for environments. He was the leading author of Gibson Env, a simulation environment for embodied perception and visual navigation, and its interactive version iGibson.
Abstract: Reinforcement Learning is considered a potential solution for acquiring many skills in different environments with minimal human supervision. However, it often struggles when the state space or the horizon of the problems grow, rendering it less useful for practical problems involving large interactive environments.
In this talk, I will present two of my recent works contributing to this topic. The first is iGibson, a novel simulation environment to develop robotic solutions for interactive tasks in large-scale realistic scenes. The scenes are replicas of 3D scanned real-world homes, aligning the distribution of objects and layout to that of the real world. Many new problems can be defined in this environment and we show evidence that solutions can be transferred to the real world.
The second is ReLMoGen, a framework to integrate motion generation into reinforcement learning. We propose to lift the action space from joint control signals to a higher level in the form of subgoals for a motion generator. By lifting the action space and by leveraging sampling-based motion planners, we can efficiently use RL to solve complex, long-horizon tasks that could not be solved with existing RL methods in the original action space.