Raw images, image labels (that determine type of object/terrain/environmental component), and point labels. Teams worked to classify each pixel as a certain category, listed in the bottom boxes, to ultimately better perception capability.
VIEW ORIGINAL
Washington— On Nov. 2, the U.S. Army Combat Capabilities Development Command (DEVCOM) Analysis Center, known as DAC, was selected as a winning team for Phase I of the 2022 Deep Green Challenge, the U.S Army’s premier data science and artificial intelligence professional development competition sponsored by the Army’s Office of Business Transformation (OBT) Enterprise Data Analytics (EDA) Office and the DEVCOM Army Research Laboratory (ARL).
DAC was one of twelve teams registered, and one of three that submitted for the final round. Working together since March, the DAC team included analysts from across the organization: Craig Andres, Matthew Banta, Ryan Barker, Paul Soper and Billy Zimmerman. In addition to the team winning overall first place, DAC received both the Best in Show award for their excellence in model development, methodology, presentation and overall performance, as well as the Most Improved Model award.
Following the 2021 challenge of predicting the materiel readiness of the M2A3 Bradley fighting vehicle, this year’s challenge focused on the perception capabilities of ground autonomous vehicle navigation systems in a wide range of environments— from off-road rocky terrain to grassy plains. It is crucial for autonomous vehicles to be able to perform well in new environments, to ultimately make intelligent decisions regarding navigation. To address this, ARL collected and annotated data from several different environments and competition teams used computer vision techniques to train and validate models to classify real-world objects and terrains. Teams worked toward a comprehensive classification model capable of categorizing every pixel in a computer vision digital image—a process called “semantic segmentation.” This challenge applied the semantic segmentation process to determine categories like fence, shrub or asphalt.
The DAC team used nine architectures and 30 encoders to create 94 distinct models to distinguish between terrain components. Of these 94 models, 74 were submitted to the leaderboard. The best six were submitted for the final evaluation.
Team Lead Matthew Banta attribute’s DAC win to their persistence. “Everyone on the team did a really great job. It was a harder problem than we expected, but we didn’t give up— we kept at it,” said Banta. “We used the leaderboard as a good indicator of how our models would do in the final and took various courses and trainings to better understand the problem and solution.”
The Army Artificial Intelligence Innovation Institute hosted the data and computing environment, and ARL’s Artificial Intelligence for Maneuver and Mobility Essential Research Program functioned as the neutral third party to evaluate models based on quantitative metrics. The Artificial Intelligence for Maneuver and Mobility Essential Research Program also hosted office hours and “Professor’s Corners” for competing teams to leverage to ask questions. According to analyst Ryan Barker, these informative sessions greatly helped guide the team’s understanding, coupled with DAC’s own external research and problem-solving.
According to Mr. Bakari Dale, the Army’s Senior Advisor for Enterprise Data Science & Artificial Intelligence, the Deep Green Challenge is an opportunity for OBT to partner with Army organizations focused on cutting-edge science and technology. “The competition is a unique opportunity for our Army’s data scientists, engineers, statisticians, and operations research analysts. Our best and brightest have been challenged to tap into their potential and develop innovative data driven solutions that can enable Army autonomous vehicle terrain perception.”
The focus for the next phase of the challenge will be on multi-modal fusion: the ability for autonomous systems to fuse information from different types of sensors, such as combining color camera data with depth data from LiDAR sensors. Phase 2 of the challenge officially kicks-off in January.
Winning algorithms will be integrated into ARL’s autonomous software framework and be further tested on live Army autonomous ground vehicles. The impact and lessons learned will be presented to the Army Analytics Board and shared throughout various forums within the broader Army and Defense communities.
By sharing the impact and the lessons learned, the pool of participants could expand from DOD-only to non-government institutions, from academia and industry that team up with a DOD sponsor. The expectation is that this larger pool will lead to more diverse and innovative approaches to address any and all challenges.
DAC is one of eight science and technology centers of the U.S. Army Combat Capabilities Development Command (DEVCOM), an organization dedicated to discovering and developing capabilities that Soldiers need to deter, and when necessary, defeat current and future adversaries. DEVCOM is a major subordinate command of Army Futures Command.
Social Sharing