Computational Clusters for Robotic Deep Learning in Complex Spatiotemporal Environment
Abstract
This proposal seeks funds for the acquisition of a GPU/CPU deep learning cluster for processing real-time sensor data in complex spatiotemporal environments. The cluster will support existing robotic facilities including UCSD Aerodrome and UCSD Autonomous Delivery Vehicles and integrate with UCSD Nautilus, a distributed cluster based on Google s open-source Kubernetes. The equipment aims to close the loop of robotic learning from data acquisition, deep learning, to automatic control in unmanned vehicles. Recent technological advances in embedded sensing and computation have enabled massive real-time sensor data for unmanned ground vehicles (UGVs) and unmanned aerial vehicles (UAV). For example, the 2 Mail Delivery Vehicles alone with 8 cameras at co-PI Christensen s generate 2 Terabytes of data per day [1]. The UAVs and UGVs in co-PI Atanasov s lab (Fig. 3, Fig. 4) are equipped with various sensors (Fig. 5), generating spatiotemporal data including thermal, RGB, and depth images and inertial and range measurements. In addition to the raw sensor data, there is also processed data generated by the computer vision and simultaneous localization and mapping (SLAM) algorithms (Fig. 1). Furthermore, high-fidelity simulators such as CARLA [2] also produce considerable amount of simulation data for various scenes. Therefore, there is an urgent need for a high-performance GPU/CPU cluster to process and analyze such spatiotemporal data. The proposed equipment will have a significant impact on PI Yu s current ARO grant# W9 l 1NF-20-1-0334: "Physics-Guided Learning for Sample Efficient Spatiotemporal Decision Making", as well as the work by co-Pis Christensen and Atanasov s participation in the ARL Distributed and Collaborative Intelligent Systems and Technology (DCIST), Collaborative Research Alliance (CRA), and many areas of interests to DoD. The cluster will also serve as a central computational platform for the education and training in robotics learning, serving 50+ faculty, 100+ Ph.D students, and 200+ Masters students affiliated with UCSD s Contextual Robotic Institute (CRI).
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 02, 2022
- Source ID
- W911NF2210179
Entities
People
- Qi Yu
Organizations
- Army Contracting Command
- United States Army
- University of California, San Diego