Sept 29, 2021 - Dongfang Liu - Video detection, segmentation, and navigation for autonomous agent

Abstract:
Autonomous agents (AA) have garnered a flurry of research interest in recent years, as they hold valuable potential for various real-world applications (e.g., self-driving cars and delivery robots). Aside from measures from fancy sensors (i.e., radar and lidar), one central objective for AA to achieve autonomy is to have the capacity of video detection and segmentation to understand the underlying environment. For detection, the challenge is compounded by the daunting issue of feature degrading intrinsic to video motion, presenting a roadblock towards developing a perception system of performance guarantees. For segmentation, the difficulty of accurately depicting instance boundary exacerbates the challenge of achieving a fine-grained representation. Consequently, developing a novel algorithm to address the above challenges inevitably lies at the core of cutting-edge research for computer vision and robotics. In addition, another core objective for AA is mapless navigation through vision without relying on odometry, GPS, or prior knowledge. With the recent progress of deep learning, a line of concurrent work has achieved advances by using deep reinforcement learning (DRL) in visual navigation. However, the DRL algorithmic underpinnings of vision-based navigation face two outstanding challenges: (1) primitive training strategy which is prone to produce suboptimal policy; (2) awkward transferability capacity which struggles to generalize to unseen scenes or domains. Therefore, understanding and obtaining a more refined picture of a vision-based navigation algorithm is the key enabler for future advances for autonomous agents and other autonomous systems in general.

Bio: