Prithwiraj Paul | Roboticist

Hi, I'm Prithwiraj Paul

Competent in

Dynamic Robotics Software Engineer and emerging leader, with a zeal for challenging boundaries in Robotics and a rapid, enthusiastic learning approach

About

As a dedicated Robotics Software Developer with a robust background in both Mechanical and Computer Engineering, I specialize in advancing the frontiers of Motion Planning, Safe Optimal Control, and Visual SLAM. My technical proficiency spans Python, C++, MATLAB, and ROS, underpinned by a strong adherence to software engineering best practices. My journey has led me through significant roles at the Indian Space Research Organization (ISRO), where I gained extensive experience in spacecraft dynamics, designing control systems, and astronaut training simulators. At Symbotic, I made impactful strides in warehouse operational efficiency by developing and implementing innovative real-time trajectory generation algorithms. My work at UC San Diego's Existential Robotics Lab further sharpened my skills and knowledge with NVidia embedded computers, where I played a key role in building a fully autonomous software stack (Planning, Localization and Controls) for a robotic car, incorporating extensive hardware testing and deployment.

I am well-versed in the nuances of Robot Testing, ensuring that my contributions not only meet but exceed performance benchmarks. My expertise extends to applying sophisticated software engineering practices in real-world robotics applications, seamlessly integrating theory with practice. I am passionate about applying my comprehensive skill set in ROS, Deep Learning, and Autonomous Systems design in a challenging full-time role, aiming to drive innovation and efficiency in the dynamic field of robotics.

Experience

Symbotic

Bot Controls Software Engineering Intern

Developed Real-time Safe Trajectory Planning Software for SymBot - Line Follower Warehouse Automation Robot
Utilized SimbaSim Gazebo for visualization and analysis of offline turn profiles, gaining insights into bot-driving behavior
Implemented Model Predictive Control for real-time trajectory generation - achieved higher convergence error and longer planning time (> 1 s) Python Libraries used: CasADi, Numpy, IPOPT Solver
Executed CBF (Control Barrier Function) - CLF (Control Lyapunov Function) framework to generate real-time safe trajectory through a faster and more efficient Quadratic Programming (QP) formulation (reduced planning time from > 1 s to < 0.3 s)
Validated trajectories, achieving over 46% optimization time improvement and 10% reduced turn time versus the baseline method
Successfully integrated online turn planning API into Symbot’s path planner software, conducted software development in C++
Implemented and executed rigorous unit tests using Google Test, ensuring robust functionality and performance improvement
Experienced in working as part of an Agile Scrum team and used Jira for bug tracking and task management and GitLab for software version control
Tools: C++, Python, GitLab, Gazebo, Jira, GoogleTest

June 2023 - December 2023 | Wilmington, M.A, U.S.A

Indian Space Research Organization (ISRO)

Space Scientist, Human Factors Engineering Entity

PROJECT 1: Designed active control system that alleviated water impact loads on Crew Module during parachute descent phase of atmospheric re-entry
Developed a 6-DOF trajectory simulator in Simulink for a 5-tonne crewed spacecraft simulating the atmospheric re-entry dynamics
Successfully implemented a robust Bang Bang controller with Proportional Derivative control action with 100% duty cycle
Performed sensitivity studies and achieved a minimum propellant mass of 22 kg (cold gas thruster) for 450 N control force
Obtained results show successful execution of Bang-Bang control to achieve the desired orientation at touchdown but call for a high control force and thrust requirement by running in 100 % duty cycle (simulation done in Simulink and MATLAB)
PROJECT 2: Designed and Developed Astronaut Training Simulators for India’s first manned mission, Gaganyaan
Configured the Static Mock-up and VR simulator defining the end-to-end requirements between the hardware and software modules
Realized an on-board computation tool conveying the spacecraft’s attitude in LVLH reference frame for crew situational awareness
Developed an autonomous quick touchdown point prediction tool during on-orbit spaceflight achieving a prediction error of < 1%
Tools: MATLAB, Simulink

September 2020 - August 2022 | Bangalore, India

BOSCH

Mechanical Engineer Intern, Automotive Aftermarket Division

Worked on hardware design changes to improve efficiency of vehicle RAC machine in the Bosch Automotive Aftermarket division
Tools: MATLAB, Simulink, SolidWorks(CAD), ANSYS - Structural

May 2018 - July 2018 | Bangalore, India

Projects

                
Autonomous NavigationDetails >
                  Vision-guided Autonomous Navigation using Reinforcement Learning
                
AccomplishmentsSkills: Robot Perception, Deep Neural Network, Transformer, CNN, ROS, PyBullet, Hardware-Testing and Deployment, OpenCV
Developed a Neural Network model for learning autonomous navigation in Ackermann drive vehicles, achieving over 95% accuracy with a
                     reinforcement learning agent trained in a PyBullet and Stable-Baselines3 simulated environment
Executing Sim2Real pipeline for autonomous navigation - Deployed trained RL model using custom C++ ROS node in Nvidia TX2, 
                    and benchmarked neural network architectures such as CNN, Transformer, Vision Transformer for image feature analysis

Online Camera CalibrationDetails >
                  Image Jacobian for Improvement of Online Camera Calibration
                
AccomplishmentsSkills: Computer Vision, Deep Learning, Pose Estimation, PyBullet, OpenGL, OpenCV, Roboticstoolbox-Python, PyTorch
Streamlined camera calibration with Image Jacobian and DREAM Neural Network for pose estimation, and executed PnP
                    with Forward Kinematics using DREAM, camera intrinsics, and joint angles for enhanced camera pose accuracy
Utilized Image Jacobian for joint position mapping in virtual camera space with OpenGL for a 7-DOF Panda Robot,
                     and designed a Proportional controller reducing camera angle error by >90% and position error by >95%

Visual ServoingDetails >
                  Image-based Visual Servoing – Tracking a moving object
                
AccomplishmentsSkills: Computer Vision, Deep Learning, Pose Estimation, PyBullet, OpenGL, OpenCV, Roboticstoolbox-Python, PyTorch
Using Image Jacobian to get the desired end-effector position
                  
Proportional controller and Inverse Kinematics to track the desired position
                  
Stochastic Optimal ControlDetails >
                  Performance Evaluation of Trajectory Tracking Algorithms for Mobile Robots
                
AccomplishmentsSkills: Model Predictive Control (MPC), Global Policy Iteration (GPI), Non-linear Control, Python, CasADi
Formulated Markov Decision Process; defined Value function, State Space, Control Space, Dynamic and Kinematic constraints
Evaluated tracking performance of RHCEC (Receding Horizon Certainty Equivalent Control) and GPI (Global Policy Iteration)
CEC: Transformed stochastic OC to deterministic Finite Horizon OC, solved as Non-Linear Program using CasADi solver
GPI: Approximated continuous-time system as discrete-time system, used Value Iteration algorithm for optimal control policy
Developed a multi-threaded 3-tier Skills-Tactics-Plays architecture for controlling omni-directional robots using ROS
Results showed CEC is Robust in real-time while GPI is offline and slower, less noise-sensitive, and discretization dependent

Motion planning in 3D obstacled environmentsDetails >
                  Benchmarking Sampling-based and Search-based Motion Planning algorithms in 3D environments
                
                Accomplishments
                Skills: 3D Motion Planning, A*, RRT*, Heuristic Function, 3D Graph-search, Data Structures and Algorithms, Python
Implemented a collision-checking mechanism for a robot’s safe navigation in 3D maze-like environments towards the goal
Implemented and assessed weighted-A* and RRT, RRT* algorithms for the robot’s goal-reaching performance
RRT achieved 30% faster and more memory-efficient performance, while A* showed superior path quality with shorter path
 Provided insights into expanded nodes, sampling method heuristic selection (Euclidean, Manhattan distance), aiding algorithm 
                    selection based on complexity, efficiency trade-offs, and graph creation efficiency in the sampling-based approach

Routing Engine development using Open Street MapsDetails >
                  Navigation API development using Open Street Maps
                
AccomplishmentsSkills: Motion Planning, A* algorithm, Data Structures and Algorithms, C++, Unit Testing, Software Development, CMake 
Crafted and deployed a C++ navigation API using OpenStreetMap data, significantly enhancing real-time route planning
Enhanced navigation accuracy and performance by implementing the A* algorithm through effective API integration 
Conducted extensive unit testing with Google Test for high-precision vehicle positioning using RTK GPS system integration

Indoor SLAM - Autonomous CarDetails >
                  ROS-Based Indoor SLAM and Mapping on Nvidia Xavier Autonomous Racecar
                
AccomplishmentsSkills: ROS, Hardware Integration and Testing, Sensor Fusion, SLAM, Point Cloud, OctoMap, Voxel Mapping, C++, Gazebo
Used ROS, Intel RealSense, and Hokuyo 2D Lidar: Conducted hardware calibration and testing; verified functionality in Rvi
Utilized Hector mapping with lidar data and generated real-time 2D occupancy grids; created octomaps using Octomap ROS package
                    and point cloud and voxel maps using Voxblox , enhancing sensor fusion and data integration
Utilized Gazebo Car Models and RealSense: Enabled detailed testing and texture mapping for efficient robot path planning

Visual Inertial SLAMDetails >
                  Visual Inertial SLAM using Extended Kalman Filter (EKF)
                
AccomplishmentsSkills: VSLAM, Kalman Filter, Sensor Fusion, IMU and Camera Calibration, SE(3) Kinematics, Map Accuracy, OpenCV 
Conducted sensor fusion for data synchronization between IMU and stereo camera, implementing EKF for real-time positioning 
                    and orientation updates of autonomous car pose using SE(3) kinematics and landmark locations
Analyzed motion and observation model noise sensitivities, achieving 95% environment mapping accuracy
                     through simultaneous car pose and landmark correction using observation model Jacobians

Particle Filter SLAMDetails >
                  LiDAR-based Particle Filter SLAM
                
AccomplishmentsSkills: Indoor SLAM, IMU and Camera Calibration, Sensor Fusion, 2D Occupancy-grid Map, Texture Map, Python 
Executed LiDAR-based Particle Filter SLAM with Magic Robot, attaining < 5cm localization error and >95% map accuracy
Crafted motion model with Encoder, IMU, Gaussian noise, scan-grid correlation for precise robot pose, and occupancy-grid mapping
Enhanced mapping and localization with Stratified Resampling, utilized Kinect RGBD camera for accurate texture mapping

Fast Class-Based Neural Style TransferDetails >
                  Fast Class-Based Neural Style Transfer (NST)
                
AccomplishmentsSkills: Computer Vision, Deep Learning, PyTorch, Semantic Segmentation, RCNN, Encoder-Decoder, Neural Style Transfer (NST)
Developed Real-Time Style Transfer Pipeline: Achieved 25 fps on Tesla T4 GPU using a custom 1.64MB Fast NST model
Boosted Semantic Segmentation with Fast-SCNN: Achieved 68.0% Mean IoU on Cityscape at 123.5 fps, refining style transfer
Optimized Style Transfer Using pre-trained VGG16: Implemented improved content and style loss (Gram matrix) 
                    for qualitative style transfer studies on CityScape and PascalVOC datasets

Markov Decision ProcessDetails >
                  Autonomous Navigation in a Door & Key environment
                
AccomplishmentsSkills: Python, Gym AI (minigrid)
Dynamic Programming: Utilizes the Dynamic Programming approach to solve motion planning problems in an environment with doors and keys
Markov Decision Process (MDP): Models the environment as an MDP, including state space, control space, state transitions, and associated costs 
Value Iteration Algorithm: Implements the Value Iteration algorithm to compute an optimal policy for navigation

3D Orientation Tracking of Rigid BodyDetails >
                  Quaternion Trajectory Tracking of Rotating body using IMU measurements 
                
AccomplishmentsSkills: Machine Learning, Trajectory Optimization, IMU Calibration, Sensor Fusion, JAX, JAXNumpy, transforms3D, Panorama
Conducted IMU calibration, including raw ADC values conversion and bias correction, ensuring accurate sensor measurements
Implemented Machine Learning model with Projected Gradient Descent, optimizing quaternion trajectory using IMU data
Constructed a Panoramic image by stitching the RGB camera images over time based on the optimized quaternion trajectory

Phone DetectorDetails >
                  Phone Detection Pipeline using Deep Learning (Object Detection)
                
AccomplishmentsSkills: Computer Vision, Deep Learning, Machine Learning, PyTorch, OpenCV, RCNN, Git, Version-Control 
Developed a PyTorch and OpenCV-based phone detection pipeline, achieving >95% accuracy on a labeled dataset using 
                    a customized Faster R-CNN model with ResNet-50 backbone and FastRCNNPredictor
Performed mini-batch training, with 0.005 learning rate, 0.9 momentum, weight decay, SGD optimizer, and GPU acceleration
Achieved >90% detection accuracy on the test dataset by engineering a custom dataset class for efficient data management, 
                    including image resizing, normalization, and bounding box extraction for precise phone localization

Teaching

Electrical and Computer Engineering, UC San Diego
- Graduate Teaching Assistant - ECE 276A, Sensing and Estimation in Robotics (Winter 2024)
- Facilitating student learning in Bayesian filtering, SLAM, Convex Optimization – taught skills in probability, linear algebra
- Guiding ~140 graduate students in projects and homework focusing on SLAM, Sensor Fusion, Sensor Calibration, and CV
Mechanical Engineering, IIT Kharagpur
- Undergraduate Teaching Assistant - Machine Design (Autumn 2020)