Learning through Interaction: Generalization in Robot Reinforcement Learning

Speaker: Chelsea Finn , UC Berkeley

Date: Wednesday, April 19, 2017

Time: 11:00 AM to 12:00 PM Note: all times are in the Eastern Time Zone

Public: Yes

Location: Seminar Room G449 (Patil/Kiva)

Event Type:

Room Description:

Host: Josh Tenenbaum, MIT BCS/CSAIL

Contact: Jiajun Wu, jiajunwu@csail.mit.edu

Relevant URL: https://people.eecs.berkeley.edu/~cbfinn/

Speaker URL: None

Speaker Photo:
None

Reminders to: seminars@csail.mit.edu

Reminder Subject: TALK: Chelsea Finn: Learning through Interaction: Generalization in Robot Reinforcement Learning

Deep learning for computer vision, speech, and natural language has demonstrated great performance in diverse real-world settings; and in my prior work, we have shown that it is possible to acquire deep neural network control policies using reinforcement learning, for robotic manipulation skills in the real world. Yet, I argue that we have yet to see the real generalization benefits of deep learning in robotics and reinforcement learning, because the established and well-understood paradigm is to learn a single task in a single environment, starting from scratch. In this talk, I will discuss two recent approaches for learning generalizable and adaptable behavior. In the first approach, we show that a robot can learn zero-shot generalization for interacting with new objects, by learning a general-purpose visual model of the world from data collected by several robots and hundreds of objects. In the second, I will present a simple meta-learning method that can achieve few-shot adaptation to new simulated locomotion tasks, and can also be applied to few-shot image recognition. To conclude, I will discuss three future research directions that I believe will be crucial for learning behavior in the real-world that can adapt and generalize to new tasks and environments, a key aspect of intelligence.

Bio: Chelsea Finn is a PhD student in Computer Science at UC Berkeley, studying machine learning for perception and control of embodied systems. She is interested in how learning algorithms can enable robots to autonomously acquire complex sensorimotor skills. During her PhD, she has developed methods for concurrent learning of perception and control for robotic manipulation, learning nonlinear cost functions through inverse reinforcement learning, and meta-learning for fast, few-shot adaptation of behavior. Chelsea received her Bachelors degree in Electrical Engineering and Computer Science at MIT. She has also spent time as an intern at Google Brain, working on self-supervised robot learning algorithms using predictive models with data from several robot arms. Her graduate research has been supported by an NSF graduate fellowship. With Sergey Levine and John Schulman, she is also currently designing and teaching a course on deep reinforcement learning, with over a thousand followers online.

Research Areas:

Impact Areas:

This event is not part of a series.

Created by Jiajun Wu at Tuesday, April 11, 2017 at 10:34 PM.