- Building the next generation of software and sensor infrastructure to evaluate the safety performance of autonomous vehicles.

◦ Implemented AI and machine learning algorithms in computer vision for automating welding processes inmanufacturing.◦ Contributed to three scalable products for assembly and welding operations.◦ Utilized deep learning platforms for 3D object structure perception, addressing computational geometry challenges.◦ Maintained and improved large-scale software in production, achieving a 5 percent accuracy improvement with a 9000 image dataset. Deployed models trained with Tensorflow and Pytorch with CUDA on nvidia GPU’s.◦ Deployed software for benchmark robotics problems using Docker images and GPU computational resources.

◦ Worked on optimization algorithms for pose estimation, SLAM, single or multiple sensor extrinsic calibration, and structure from motion using total least squares.◦ Enhanced position (up to 1 cm 3-sigma bound) and orientation accuracy (up to 2 degree 3-sigma bound) of large-scale camera and 3D sensor datasets.◦ Developed sensor fusion algorithms to increase reliability in case of missed data samples from one sensor, in the context of multi-sensor configurations. Benchmark problems are camera calibration, structure from motion, point cloud and image registration.I am mostly working on the experiments related to my thesis here. We have a multi-camera tracking system as a ground truth for our experiments. LIDARs and cameras are mainly used to verify the estimation algorithms. The main idea is a framework for the pose estimation problem using total least squares for vector observations from landmark features. First, the optimization framework is formulated for the pose estimation problem with observation vectors extracted from point cloud or image features. Then, error-covariance expressions are derived. The attitude and position solutions obtained and are proven to be optimal within the first order approximation. The proposed solution is simulated in a Monte-Carlo Simulation, Gazebo Simulation, and experiments. I am currently writing my second paper which is based on similar ideas for monocular camera pose estimation.

1-Total least Squares for Optimal Pose Estimation:This work provides a theoretical framework for the pose estimation problem using total least squares for vector observations from landmark features. First, the optimization framework is formulated for the pose estimation problem with observation vectors extracted from point cloud features. Then, error-covariance expressions are derived. The attitude and position solutions obtained via the derived optimization framework are proven to reach the bounds defined by the Cramér-Rao lower bound under the small angle approximation of attitude errors. The measurement data for the simulation of this problem is provided through a series of vector observation scans, and a fully populated observation noise-covariance matrix is assumed as the weight in the cost function to cover for the most general case of the sensor uncertainty. The proposed solution is simulated in a Monte-Carlo framework with 10,000 samples to validate the error-covariance analysis.You can view our latest paper on this topic:https://arxiv.org/pdf/2106.11522.pdf2-MMAE for noise covariance estimation:The process noise covariance for the motion of a differential drive robot is estimated through a bank of parallel filters called the multiple model adaptive estimation which provides a real time parallel-structured estimation framework for the states as well as other unknown parameters in a nonlinear system.3-Application of Large Deviations Theory in Optimal and Stochastic control:Trying to minimize the asymptotic probability of sample path deviating more than a threshold K from the nominal path as the cost function. This probability is described by exp(−n I(x)) based on large deviation theory. We are trying to develop a control policy to minimize this deviation asymptotic probability alongside keeping the deterministic quadratic cost of the system at a desired level.Software skills: MATLAB, ROS, C++, Python

TA for Space Dynamics and Control, Continuous Control, Systems Dynamics, Flight Dynamics, and Engineering Drawing. Responsible for grading theoretical and MATLAB homeworks and teaching the concepts of these courses in the recitations and office hours.

Title: A Hamiltonian Based Approach to solve the Fokker Plank Kolmogorov Equation for a Class of Nonlinear Systems

◦ Implemented deep learning platforms for object detection, improving picking accuracy by 2 percent by deploying a post processing algorithm on top of the main segmentation pipeline to better identify the geometry of the objects in 2D images and therefore a more efficient pick and drop functionality of the robot.◦ Tuned networks on real datasets with at least 20,000 images of warehouse packages.◦ Optimized extrinsic calibration of RGB and RGBD cameras.◦ Gained experience in bash scripting in Linux.The main task was investigating the showcase problem of picking a series of objects using visual data. After pre-processing the data and going through the deep learning framework of the object detection, the main objective will be how to identify and extract the geometry of a given object, using the image processing algorithms. The application was a loader-unloader robot inside of a trailer using the main perception setup of LIDARs and cameras.I also had a literature review on robot calibration, about the hardware required to deliver the minimum possible errors and also the mathematical techniques to do an optimization of a cost function of errors. This task is generally called homing and calibration of the robot manipulators and arms in the literature. The other task involved a research on the methods we can use to fuse the information from different type of measurement units like LIDARs and pose sensors. I was skimming through a series of github repositories and papers to find a computational efficient as well as accurate way of information fusion.Another problem that I had research about was how can we compensate for an object that is missed during the detection step of one sensor using the measurement information we get from the other sensors. For example if we miss something from the camera object detection, can we robustly compensate for the position of the missed object using a LIDAR?

◦ Developed OOP code for multi-target multi-sensor tracking, achieving a 3 percent improvement in position andvelocity estimations.◦ Implemented state estimation and sensor fusion algorithms for linear and nonlinear systems.◦ Collaborated in a diverse team with backgrounds in mathematics, firmware, and software engineering.I generally tried to contribute to the theoretical ideas and software architecture of the robotic perception team for an autonomous driving application. A specific was investigating the problem of tracking the objects which have nonlinear motion model and/or sensor measurement model. After pre-processing the data, we fuse the information from different sources and end up with a single de-noised estimation of the states to be fed to the control policy calculations. Basically we had a multi-target multi-sensor tracking problem. We had a series of classes for the pre-processing and making a system-wide data structure for sensor, and then another series of classes for data association and assigning a tracklette to each object of interest. In the estimation pipeline, we could use several nonlinear estimation algorithms such as EKF, UKF and particle filters. EKF provides a linearization of equation motion and sensor models and forwards the provided matrices to a Kalman filter framework that outputs a refined state estimation needed in control policy. I developed the mathematical formulation for the data association and multiple sensor fusion using the UKF as the core estimator. After developing the math and checking the code in unit tests, we integrated the code the main perception framework and saw consistent and robust results in the tracking experiments in a unity-based simulation environment and rviz in ROS. Finally, we got the developed software working on a series of collected data sets from the outdoor driving scenarios in the parking lots, intersections and streets.

Graph Embedding Algorithms in data mining with the goal of drug re-purposing. Ref. to http://compbio.buffalo.edu/research.html for more information.We have a graph of drugs and diseases(2 type of nodes). Each edge label shows the association level of a drug and disease (training data).After extracting a set of features from this network based on a cost function(could be the one implemented in node2vec algorithm) from the association information in the edges.After that each node has a feature vector that can be used for the association prediction in the test data based on a distance measure(could be a sigmoid function combined with a euclidean distance) and then we use a ROC curve for showcasing the accuracy of the embedding algorithm. 10 fold cross-validation is used for accuracy generation.Software: python