Ekta Samani
I am a Postdoctoral Scientist at Amazon Robotics. I completed my Ph.D. in Mechanical Engineering from the University of Washington (UW), advised by Prof. Ashis Banerjee. Before joining the Ph.D. program, I was an Operations Officer at Hindustan Petroleum Corp. Ltd. in Mangalore, India. I obtained my B.Tech in Electrical Engineering with a minor in Computer Science and Engineering from the Indian Institute of Technology (IIT), Gandhinagar.
The central goal of my research is to achieve human-like performance in autonomous systems by combining model-based reasoning with advanced pattern recognition. My dissertation research focused on achieving perception robustness in low-cost robots using topological representations in pattern recognition-based frameworks while drawing inspiration from the human perceptual process.
News
Jun 2024: I won the UW Graduate School's 2024 Distinguished Dissertation Award in the Mathematics, Physical Sciences & Engineering category
Jan 2024: I started serving as an Associate Editor for the IEEE Robotics and Automation Letters
Aug 2023: Joined Amazon as a Postdoctoral Scientist under the Postdoctoral Science Program
Aug 2023: Successfully completed my Ph.D. [Dissertation]
Jul 2023: UW Mechanical Engineering featured a story about my research on the department website. Check another related story here.
May 2023: I received Women Engineers (WE) Rise Outstanding Student Award (Mechanical Engineering)
Apr 2023: I have been selected as an RSS Pioneer (2023).
Mar 2023: I won second place prize in the UW Three Minute Thesis (3MT) Competition
Recent Work
We propose a first-of-its-kind human reasoning-inspired framework, termed THOR, for object recognition in unseen cluttered environments. THOR uses persistent homology (from computational topology) in a novel slicing-based approach to obtain structural descriptions of object point clouds. Systematic evaluations on real-world scenarios with different environmental conditions and degrees of object occlusion on a benchmark RGB-D dataset show better performance than state-of-the-art methods.
We introduce an effective baseline approach, F2BEV (i.e., Fisheye to BEV), incorporating fisheye projection and distortion models into a transformer-based BEV generation approach originally designed for PV images. We propose single-task and multi-task variants of F2BEV for individual and simultaneous generation of BEV height and segmentation maps from fisheye images, respectively.