ABOUT ME

I am an Assistant Research Scientist in the Engineering Systems Group at UMTRI. The theme of my research has been to address robustness and uncertainty issues in modelling high-dimensional LiDAR sensing data during my doctoral studies and in the autonomous vehicle (AV) pipeline subsequently.

During my doctoral  research, I looked at several facets of uncertainty quantification such as, in (Glennie, Kusari, & Facchin, 2016), we published the first and only detailed calibration and accuracy study for Velodyne VLP-16 laser scanner and in (Glennie, et al., 2014), we corrected legacy LiDAR datasets to remove a majority of  the systematic errors. We also developed a novel method to extract building features in earthquake zones to estimate the near-field movements during earthquakes and tested our method on the 2014 Napa valley earthquake with very accurate results (Kusari, Glennie, Brooks, & Ericksen, 2018). 

Since my stint at Ford Motor Company, I have shifted my focus towards the AV pipeline and the robustness issues surrounding the deployment. The AV pipeline can be divided into sensing, localization, perception, decision-making and control modules respectively. I have contributed to all but the control module. I have also been working on the development of the simulation pipeline responsible for the testing and validation of these modules.  

Sensing: Sensing utilises the given suite of sensors to collect data at any given time step of the environment. Recently, we published a study documenting the closed form computation object-level uncertainty computation from infrastructure mounted cameras in (Kusari, Almutairi, Gilbert & LeBlanc, 2023). Also as part of the National Highway Transportation Safety Administration (NHTSA) funded research project, we have constructed a calibration test site for joint calibration of all mounted sensors on a vehicle during motion. It consists of three targets - each consisting of three perpendicularly placed April tag corner cube reflectors - placed equilaterally with enough space to have the vehicles driven by it. We have also devised a methodology to accurately calibrate the sensors using a graphical approach that is awaiting patent procedure. My future research directions are in providing convergence guarantees in multi-sensor calibration and understanding characteristics of weather conditions for sensing data. 

Localization: Localization is used to infer the position of the AV in the world. At Ford, my first responsibility was to create a scalable mapping pipeline from LiDAR data. In my current capacity, I have been a co-PI of a NHTSA funded project on testing ADS localization systems using different modalities such as a separate LiDAR sensor placed on top of the vehicle and gathering data for fiducial cylindrical targets in a closed testing course. As part of the project, we created the operational requirements for ADS systems using a Weibull distribution and provided different scenarios such as highway, urban, turning at intersections etc (Kusari & Saha, 2023). I plan on developing scalable localization algorithms. 

Perception: Perception is used to convert sensing data into objects based on features such as appearance, shape and roughness. Currently, perception is completely driven by model-free deep learning methods which can achieve human-level accuracy. However, deep learning suffers from limitations which can cause the system to fail in unexpected and fatal ways. One significant limitation is the failure in out-of-distribution (OOD) cases where a trained neural network when shown data from an unseen class gives completely wrong answers. Currently all OOD detection methods use some samples to construct the separation boundary between the in-distribution (InD) and OOD. In (Chen, Sung, Kusari & Sun, 2024) we develop a novel Gaussian process (GP) based OOD detection method which constructs boundaries based on only InD data without seeing any OOD samples. We also provide theoretical justifications and show benchmarking results on popular open-source datasets. We are also starting to investigate the effect of sensor failures on perception, the idea being that sensor failures are unpredictable and the resulting samples constitute OOD (Prabhakar, Girnar & Kusari, 2024). Our goal is to measure the failure caused by sensor errors in a robust fashion in the real world. For LiDAR perception, we have devised a novel graph-theoretical approach to infer normals for LiDAR data (Kusari & Sun, 2022). My future plan would be to perform minimalist LiDAR perception rather than the large compute models that are the state-of-the-art currently. 

Decision-making: Decision-making takes in the dynamic objects, map information and the objective to find an optimal action for the AV. Reinforcement learning (RL) has been proposed as a probabilistic paradigm to provide sequential decisions based on states. It has also shown to achieve great performance in AV decision making while accounting for the uncertainties in states. However, RL suffers from the black-box nature of neural networks which makes the results unexplainable. In (Wan, Li & Kusari, 2023), we infer explainability and causality for understanding the physics of lane changing decisions based on the neighbouring vehicle behaviour. 

Simulation: I have devoted a large part of my research towards generating better simulation procedures for training RL algorithms (Kusari, Li, Yang, Punshi, Rasulis, Bogard & LeBlanc, 2022). As part of the NHTSA project, we have also developed a large-scale scenario database of photorealistic simulation environment (CARLA) providing multi-sensor data collection with a variety of vehicles and environmental conditions. I have also written a book chapter on the state of simulation for AVs which I hope would be used by practitioners as a guide to structure high-dimensional large-scale simulation (Kusari, 2024).

Dr. Kusari was chosen as the academic partner to industry behemoths such as Deloitte, Nvidia, AWS etc, where he was tasked with grounding the possibilities of the solutions to be practical, flexible and above all, robust.