Date: January 22, 2026

Time: 1pm EST

Location: Van Leer 218

Virtual Link: Zoom

Meeting ID: 968 3730 8246

Passcode: 650576

Andrew Yarovoi
Ph.D. Student, Robotics
George W. Woodruff School of Mechanical Engineering
Georgia Institute of Technology

Committee:

Dr. Christopher R. Valenta (Advisor)

School of Electrical and Computer Engineering 

Georgia Institute of Technology

Dr. Lu Gan

Guggenheim School of Aerospace Engineering

Georgia Institute of Technology

Dr. James Hays 

School of Interactive Computing

Georgia Institute of Technology

Dr. Zsolt Kira

School of Interactive Computing

Georgia Institute of Technology

Dr. Patricio A. Vela

School of Electrical and Computer Engineering 

Georgia Institute of Technology

Abstract:  

The proposed research aims to advance 3D point cloud semantic understand by enabling effective multi-dataset training (MDT) across diverse domains and sensors. Current models trained on individual datasets often exhibit limited generalization due to restricted point cloud dataset sizes and sensor-specific biases. MDT offers a solution by providing more diverse training data from a variety of sensors and domains. However, implementing MDT for point cloud data remains challenging due to unresolved issues such as heterogeneous input features, insufficient contextual information, significant variations in scale and density, and inconsistent labeling across datasets. The proposed work aims to systematically address these challenges through architectural design. First, a modular training framework will be developed to accommodate heterogeneous sensor inputs and align synonymous class labels through an open-vocabulary approach. Second, novel expansion and contraction network blocks will be designed to efficiently generate multi-resolution features, enhancing scale and density invariance. Third, a new backbone model, termed UniMD3D, will integrate these modules and implement a unified label alignment mechanism to resolve class partitioning inconsistencies across datasets. The proposed research is expected to enable effective learning while training on mixed datasets spanning multiple domains, tasks, and sensors, leading to more generalized 3D perception systems for real-world applications.