University Events

Description: Topic: Terrain Aware Autonomous Ground Navigation in Unstructured Environments Informed by Human Demonstrations

Location: Lester W. Cory Conference Room, Science & Engineering Building (SENG), Room 213A

Zoom Conference Link: https://umassd.zoom.us/j/3753158123 Meeting ID: 375 315 8123 Passcode: 495427

Abstract:
Mobile robots equipped with the capability to perform autonomous waypoint navigation can replace humans for applications such as humanitarian assistance, nuclear cleanup, and reconnaissance. In such tasks, the robot must be able to accurately and reliably perform complex behaviors such as the ability to navigate over unstructured terrain and respond to unseen situations similar to how a human would. However, to implement complex behaviors beyond obstacle avoidance, many current approaches employ machine learning methods requiring large amounts of labeled data. While simulations can quickly generate large amounts of labeled data, the same cannot be said for real world environments, limiting adoption of mobile robots to complete the aforementioned applications. Moreover, solutions are often brittle, exhibiting poor performance when operating outside of environments beyond where they were designed or trained. Therefore, there is a need for methods which can quickly learn navigation behaviors from limited data while being able to adapt to dynamic scenarios.

To communicate both positive and negative environmental scenarios, roboticists assign rewards (or inversely, costs) to all the relevant environmental features expected. For robots that need to quickly transition between varying unstructured environments, defining rewards prior to understanding all future features the robot will encounter is often unachievable as either: (i) the robot is unable to perceive the new feature, and (ii) the numerical reward for each feature is unknown. In such scenarios, it is more effective for a human supervisor to provide examples of desired behavior than for an engineer to explicitly define it. Therefore, this dissertation focuses on a ground robot's ability to incorporate human demonstrations as a weak supervisory signal to solve each respective preceding problem by (i) obtaining a semantic perception model capable of classifying terrains present in the current environment given a sequence of unlabeled (unsupervised) images and (ii) using Bayesian inverse reinforcement learning to learn rewards associated with the terrains identified to build cost-maps online for autonomous waypoint traversal.

Advisor(s):
Dr. Lance Fiondella, Associate Professor, Department of Electrical & Computer Engineering, UMASS Dartmouth

Committee Members:
Dr. Hong Liu, Commonwealth Professor, Department of Electrical & Computer Engineering, UMASS Dartmouth; Dr. Jiawei Yuan, Associate Professor, Department of Computer Information Science, UMASS Dartmouth;
Dr. Craig T. Lennon and Dr. Maggie B. Wigness, United States Army Research Laboratory

NOTE: All ECE Graduate Students are ENCOURAGED to attend.
All interested parties are invited to attend. Open to the public.

*For further information, please contact Dr. Lance Fiondella at 508.999.8596 or via email at lfiondella@umassd.edu.