Event
Ph.D. Research Proposal Exam: Mohamed Bashir Dafaalla Elnoor
Friday, May 9, 2025
12:00 p.m.
AVW 2328
Maria Hoo
301 405 3681
mch@umd.edu
ANNOUNCEMENT: Ph.D. Research Proposal Exam
Name: Mohamed Bashir Dafaalla Elnoor
Committee:
Professor Dinesh Manocha (Chair)
Professor Pratap Tokekar
Professor Kaiqing Zhang
Date/time: Friday May 9th, 2025, 12:00 PM – 2:00 PM
Location: Via Zoom
Zoom Link: umd.zoom.us/my/dmanocha
Title: Towards Robust and Efficient Multi-modal Perception for Autonomous Navigation
Abstract:
Autonomous navigation plays a critical role across a wide range of robotic applications such as logistics, agriculture, hospitals, surveillance, and disaster response. However, when deployed in unstructured and dynamic environments, robots must overcome challenges related to uncertain terrain properties, sensor degradation, and the need for contextual reasoning. We investigate these challenges through a sequence of multimodal perception methods that integrate vision, LiDAR, proprioception, and language models to enable robust and efficient navigation. We begin with ProNav, which estimates terrain traversability using proprioceptive signals to enhance stability and predict failures across vegetated, rocky, and granular terrains. Building on this, we introduce AMCO, which combines visual and proprioceptive data into an adaptive cost mapping framework to guide gait and velocity selection in real time. While large Vision-Language Models (VLMs) offer strong semantic reasoning capabilities, their high computational cost and latency make real-time deployment on robots challenging. To address this, we propose VLM-GroNav, which refines VLM predictions by grounding them in physical interaction, which enables improved path planning across slippery and deformable terrains. Most recently, we developed Vi-LAD, which distills the reasoning abilities of large VLMs into a lightweight, attention-aware model that supports real-time, socially compliant navigation on resource-constrained robots. We deploy and validate these methods across diverse platforms, including Clearpath Husky, Boston Dynamics Spot, and Ghost Vision 60, in a range of challenging environments.
