Python Assistant

SAI JADHAV

858-***-**** adcktz@r.postjobfree.com www.linkedin.com/in/sai-jadhav-9496ok

EDUCATION

University of California San Diego Sep’18- Jun’20

MS in Electrical and Computer Engineering 3.74/4.0 Courses: Sensing and Estimation in robotics, Planning and Learning in Robotics, Reinforcement Learning, Machine Learning for Image Processing, Machine Learning Algorithms, Human-Robot Interaction TECHNICAL SKILLS

Languages: Python, C++, Pyspark, MATLAB

Tools/Libraries: PyTorch, OpenAI Gym, Numpy, Apache Spark, Databricks, OpenCV, Pandas, Scikit-Learn, Pickle RESEARCH AND WORK EXPERIENCE

Graduate Student Researcher at UCSD, Existential Robotics Laboratory (Python, MATLAB) Sep’19-Present Single/multi-robot SLAM using CLEAR Algorithm

• Merge observations from multiple robots to form a global map of the environment using Visual Inertial SLAM.

• Increased fusion accuracy by checking for cyclic consistency and solving the multi-view matching problem in SLAM using Consistent Lifting, Embedding, and Alignment Rectification (CLEAR) Algorithm. Big Data Intern at HP, Inc (Apache Spark, Databricks) Jun’19-Sep’19

• Segregated over 2TB of Windows Event Logs from around 280 HP devices into high level categories using K means clustering and elbow method by converting the Windows Event messages into feature vectors using TF-IDF.

• Built an analysis model using event correlation which predicted failures in the system based on past events that have occurred in the device using FP growth algorithm in Databricks. Teaching Assistant at UCSD (CSE 276C-Mathematics for Robotics) Sep’19-Dec’19

• Taught topics in robotics like surface detection (floors, walls) using 3D point cloud data and RANSAC, safe navigation of robots using RRT, PRM and Vornoi diagrams, image classification using PCA and LDA etc. to a graduate class of 30 students. PROJECTS

Deep Reinforcement Learning for First Person Shooting (FPS) games (Python, PyTorch, VizDoom)

• Trained an agent to play a FPS game (DOOM) using Double Deep Q learning in the VizDoom API.

• Scene understanding and detecting features like road, weapons and humans from the first-person view of the agent using a CNN. Performed complex tasks like navigation and picking up weapons using Deep Reinforcement Learning networks like DQN and DRQN. Increased the survival time of the agent by 50% as compared to the state of the art. SLAM using Particle Filter and Visual Inertial SLAM using Extended Kalman Filter (Python)

• Implemented SLAM using a Particle filter to estimate a 2-D occupancy grid map of the environment by indoor localization and occupancy grid mapping by incorporating IMU, odometry, and LIDAR (Hokuyo) measurements from a differential drive robot.

• Generated a textured map of the environment by detecting the floor area in the image by performing depth estimation on disparity images and projecting colored points from the RGBD sensor onto the occupancy grid.

• Performed Visual SLAM using an Extended Kalman Filter on synchronized measurements from an IMU and a stereo camera using the intrinsic camera calibration and the extrinsic calibration between the two sensors. Image Segmentation and Object Detection using Statistical Learning methods (OpenCV, Python)

• Performed color-based detection of barrels in occluded images under varied lighting conditions using gaussian mixture models by segmenting the image using Maximum Likelihood Estimate and generating a bounding box around the object of interest, resulting in 79% accuracy.

Remote Heart Rate Measurement using Image Processing and Deep Learning (OpenCV, PyTorch, MATLAB, Python)

• Developed an algorithm to perform remote heart rate measurement (tolerance of 8 bpm) of a person using pixel to pixel pulse extraction, spatial pruning and temporal filtering of relevant facial regions in videos.

• Performed face detection using Deep Neural Networks on video frames and classified the detected face to discriminate skin and nonskin regions using OC-SVM trained on feature descriptors created using intensity-normalized RGB and YCrCb images. Path Planning for Robots (Python, C++)

• Implemented and compared the performance of search based and sampling based planning algorithms like weighted A*, RRT*, Bidirectional RRT and RRT-connect under strict timing constraints for a set of 3D environments by varying parameters like the choice of heuristic function, the choice of sampling and steer functions, etc. Performed 55% faster than the greedy approach. Reinforcement Learning for a 2 link Robot Arm (OpenAI Gym, PyTorch, Python)

• Trained a Robot Arm with random position initialization to reach a goal position by implementing DDPG and TD3 using Actor Critic Neural Network.

• The network calculated the Q value and action policy given the pose and orientation of the robot. Human Protein Atlas Image Classification (PyTorch, Python)

• Implemented multi label classification of HTI (high-throughput fluorescence microscopy imaging) images of human cells using a CNN. The model architecture included Transfer Learning using ResNet18, Global Average Pooling and SELU.

Contact this candidate