D3RL: Distributed, Diverse Data for Robot Learning
Abstract
Data-driven approaches to robotic learning has driven progress and performance in several robotictasks like grasping, pushing, collision avoidance, and navigation. However, the scale of data collectedin robotics still pales in comparison to fields like computer vision on two key aspects: diversityand size. First, robotic data is mostly collected on a single robot in a single environment,which causes the learned models to severely overfit. Collecting diverse examples is hence crucialfor robot learning methods to work in diverse, previously unseen environments. Second, mostrobot learning methods operate on up to 100K robot data, which is orders of magnitude smaller thanwhat is needed for more dexterous manipulation tasks such as fixturing, door-opening, and laundryfolding.This necessitates the need for robots to imitate humans to reduce the burden on exploration.Through this proposal, we will create a cross-university robot cluster with 8 multi-fingered roboticmanipulators. This industrial-grade setup will enable diverse data collection on difficult manipulationproblems with a single shared software platform, which will directly accelerate research ongeneralization and imitation in robot learning. Furthermore, the proposed equipment will impactfour of our ongoing DoD supported research projects.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2021
- Source ID
- N000142112404
Entities
People
- Lerrel Pinto
Organizations
- New York University
- Office of Naval Research
- United States Navy