Robot Learning from Internet-Scale Data
Abstract
Robot Learning from Internet-Scale DataChelsea Finn, Stanford University.Total funds requested: $1,000,000Approved for Public Releas,eAbstract. Autonomous robots that can be quickly trained to operate in novel real-world, unstructuredsettings have the potential to,be tremendously useful for a range of manual, tedious,or dangerous tasks that are currently performed entirely by humans. Learning-b,ased methods arenow the dominant framework for building systems that require perception and prediction capabilities;and the most suc,cessful deployments of these systems have been fueled by pre-training on diverse,Internet-scale datasets. We propose to develop a di,verse, Internet-scale robotic manipulationdataset, along with algorithms that can leverage that data to learn general-purpose repres,entationsthat are useful for learning many different downstream tasks. This proposition, however, presentsmultiple major research ch,allenges. First, robot data does not already exist in massive quantities,and while robots can in principle collect data autonomously,, we lack algorithms that can produceuseful autonomously-collected data. Second, existing robot learning algorithms typically assume,special-purpose data (e.g. demonstrations or online data collection) or significant human involvementsuch as detailed reward supervi,sion or human interventions to reset the environment. Toaddress these challenges, we aim to perform research in three core direction,s: First, can we understandwhat data is needed for good robot generalization and develop techniques for collecting suchdata at scale, in the real world? Second, can robots learn general-purpose sensorimotor representationsin a fully self-supervised manner, that for,m useful priors over both perception and action?Finally, can we efficiently adapt general pre-trained representations to specific do,wnstream tasksusing minimal human supervision?Technical Approach. The proposed research will be organized around three main thrusts,, whichwill address different aspects of robotic learning from Internet-scale data.Part 1: Internet-Scale Data for Robotic Manipulati,on. We will study how data composition affectsrobot generalization, and techniques that allow us to scalably collect diverse robotic, manipulationdatasets that are 10-100x more diverse than prior datasets.Part 2: Self-Supervised Pre-Training for Robotic Manipulatio,n. We will study self-supervised approachesfor learning a single general-purpose pre-trained model from diverse robot data, includin,gapproaches based on visual dynamics modeling and model-free methods.Part 3: Downstream Robot Task Adaptation. We will study how a p,re-trained robotic foundationmodel can be leveraged to accelerate training and online exploration for a variety of differentdownstre,am tasks.Anticipated Outcomes. The proposed research will advance the state-of-the-art of robot learning,making it possible for robo,ts to learn a task in a novel environment with little human effort.It will also advance our understanding of how the diversity and c,omposition of a dataset affectsgeneralization in robot learning.Impact on DoD Capabilities. Robotic systems that can learn from broa,d datasets will be moreuseful in real, unstructured environments, allowing them to perform a variety of tasks includingcleaning, mai,ntenance, repair, and construction.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 13, 2022
- Source ID
- N000142212621
Entities
People
- Chelsea Finn
Organizations
- Office of Naval Research
- Stanford University
- United States Navy