Building and Benchmarking Self-Driving Capabilities in Robots
Discipline: Technology and Engineering
Subcategory: electrical/computer
Session: 1
Room: 3 - Hanover C
Johann Mission - University of Maryland Baltimore County
Co-Author(s): Meriel Von Stein, Trey Woodlief1, Mathushan Mathyvannan1,2, Yili Bai1, Zachariah Risheq1, Sebastian Elbaum, PhD1 1 University of Virginia 2 University of Maryland, Baltimore County Charlottesville, VA, USA Baltimore, MD, USA
Programmable robots with self-driving capabilities are employed for a growing number of tasks, such as food delivery, traffic safety, and security. Such systems often struggle to safely navigate complex new environments. To better understand how and why they struggle we have built a benchmark consisting of a complete autonomous navigation system and a set of target environments. Building and benchmarking this system requires substantial data collection, training, and testing. Towards achieving this goal, we collected 60 runs of data spanning 11 hours from a human driver operating a Husarion ROSbot XL robot in a series of hallways, totalling over 30,000 observations. Each observation includes images, LiDAR sensor readings, and speed. Raw data was cleaned to remove failed instances such as controller malfunctions and crashes, and programmatically augmented by flipping, saturating, and shadowing the images to double the amount and variety of useful data for successful training. We selected the DAVE2 model, trained it, and deployed it in a hallway to operate autonomously. Preliminary results after 3 hours of successful deployments show that the robot can avoid collisions while navigating hallways, as well as successfully detour around a person or obstacle. However, the sensors and trained model fail to accurately detect glass walls and thin obstacles such as table legs and cables. Planned studies will provide a more significant assessment including multiple environments. These findings and lessons learned, together with the benchmark, will enable researchers to build autonomous systems that can provide safer operations in new environments.
Funder Acknowledgement(s): This research was sponsored in part by the University of Virginia, Leadership Alliance, National Science Foundation, Defense Advanced Research Projects Agency, Air Force Research Laboratory, and Army Research Office
Faculty Advisor: Dr. Sebastian Elbaum, selbaum@virginia.edu
Role: Summer Research Intern