Title: Semantic Scene Description for Distributed Simultaneous Localization and Mapping in Communication-Constrained Environments

Date: Friday, January 27, 2023

Time: 9:00AM – 11:00AM ET

Location: TSRB Room 523

Tony Lin

Robotics PhD Student

School of Electrical and Computer Engineering

Georgia Institute of Technology

Committee:

Dr. Fumin Zhang (Co-Advisor) – School of Electrical and Computer Engineering, Georgia Institute of Technology

Dr. Samuel Coogan (Co-Advisor) – School of Electrical and Computer Engineering, Georgia Institute of Technology

Dr. Matthieu Bloch – School of Electrical and Computer Engineering, Georgia Institute of Technology

Dr. Seth Hutchinson – School of Interactive Computing, Georgia Institute of Technology

Dr. Ye Zhao – School of Mechanical Engineering, Georgia Institute of Technology

Abstract: 

This proposal aims to develop a framework to solve a distributed Simultaneous Localization and Mapping (SLAM) problem in communication-constrained environments, in which robots individually solve the SLAM problem while sharing semantic descriptions (e.g., “I see potted plants on my left and a television on my right”) of their personal views. Our proposed approach incorporates such information by treating semantic descriptions of scenes as coarse relative pose estimates perturbed by some non-Gaussian noise which must be learned from data. Central to utilizing this semantic information is handling the inherent difficulties associated with the quantifying the unknown noise distribution and the data association problem arising from the lack of uniqueness in semantic descriptions (multiple scenes may be described by the same semantic description). To overcome these difficulties, this work leverages a particle-driven filtering and smoothing strategy in which Generative Adversarial Networks (GANs) are utilized to learn the unknown noise distribution and a novel Particle Product strategy, which approximates the product of two particle distributions, is used to transform and fuse shared particle distributions into a robot’s local frame of reference. The proposed Particle Product strategy is also employed to provide resiliency in the data association problem when performing smoothing of the particle estimates for each robot’s trajectory.