Miheer Diwan

Robotics Engineer

I'm working at the intersection of Computer Vision and AI to bridge the sim-to-real gap!

I recently completed my Master's degree in Robotics Engineering from Worcester Polytechnic Institute (WPI), Worcester, USA where I worked at the PeAR Lab on Vision and Navigation projects.

I also hold a Bachelor's degree in Mechatronics Engineering from NMIMS University, Mumbai, India.

Email  /  Resume  /  LinkedIn  /  Github  /  CAD

I am actively looking for Internships/ Full-Time opportunities!

Education

Relevant Coursework:
Computer Vision, Machine Learning, Robot Controls, Hands-On Autonomous Aerial Robotics, Advanced Robot Navigation, Project Management

Affiliations:
Perception and Autonomous Research (PeAR) Group , Embedded Computing Lab , WPI Innovation Studio , Squash Club

GPA : 3.87

Relevant Coursework:
Digital Signal & Image Processing, Dynamic Systems Modeling and Analysis, Microprocessors and Microcontrollers, Robotics System Design, Design of Machine Elements

Affiliations:
Technotix Robotics Team, Team Dramatics

GPA : 3.46

Skills

Robotics: Perception, Localization, Deep Learning, ADAS, Aerial Robotics, State Estimation, Sensor Fusion, Calibration

Software: Python, C++, ROS/ Gazebo, OpenCV, PyTorch, TensorFlow, CUDA, TensorRT, NumPy, SciPy, PCL, Open3D, scikit-learn, Matplotlib, MATLAB/ Simulink, Docker, Linux, Git, LaTeX, SolidWorks, Blender, CoppeliaSim, NI LabVIEW

Hardware: DJI Tello EDU, Jetson Orin Nano, Arduino, Raspberry Pi, Yaskawa Motoman MH5, 3D Printing, Rapid Prototyping, Oak-D Pro, Realsense D435, Velodyne VLP-16 Lidar

Network Architectures: YOLO, CNN, DERT, R-CNN, Fast R-CNN, VGG16, ResNet18, DenseNet, U-Net, LSTM, TCN, HomographNet , NeRF

Languages: English (native fluency), Hindi (native fluency), Marathi (native fluency)

Experience

• Collaborated with the perception team to optimize, upgrade, and deploy the perception stack of Trossen Robotics’ UGV.
• Developed a ROS 2 package using C++, PCL, and CUDA to efficiently process, denoise, and merge LiDAR and stereo camera point clouds.
• Generated and optimized point clouds using the Oak-D Pro stereo camera to achieve precise detection of objects as small as 8 cm on the ground.

Advisor: Prof. Nitin J Sanket

Sim-2-Real Autonomous Quadrotor Navigation in the Wild

• Developed a Generative, Procedural environment for quadrotor simulation and learning using Blender software.
• Employed RRT* algorithm and Minimum Snap Trajectory generation to navigate a dense, simulated forest.
• Designed a Model Predictive Control for precise trajectory tracking and traversal of optimal trajectories.

Advisor: Prof. Xinming Huang

Voxel-based Autonomous Driving and Point Cloud Segmentation

• Developed the perception stack for autonomous vehicles with camera images, LiDAR point clouds and Voxel Grids.
• Estimated absolute and relative depth using both Stereo Photogrammetry and Intel’s MiDaS depth model.
• Used YOLOv8 for Object Detection and Instance Segmentation for road vehicles, traffic signs, and pedestrians.
• Detected lanes on custom dataset with a 97.8 % accuracy using Cross Layer Refinement Network (CLRNet).

Back-EMF-based Sensorless Control of BLDC Motor Using MATLAB/Simulink

- Conducted comprehensive research, exploring and analyzing multiple non-linear methods for BLDC motor speed control.
- Constructed a closed-loop model for sensorless control of BLDC motors using Back-EMF and zero-crossing detection.

- Headed a team of 15 people to develop four mobile robots: Arrow-Shooting robot and Ball-Throwing robot.
- Fabricated a custom swerve drive mechanism for the robots which provided 30 % increase in the speed of the robots.
- Secured India National Finalist positions in ABU Robocon 2020 and 2021, with a flawless score in the Design Phase.

Computer Vision & Deep Learning

Per-Pixel Probabilistic Boundary Detection Algorithm

Stitched images having overlapping features to create a panorama using feature matching, ANMS, homography estimation, RANSAC.

This project leverages classical computer vision and Deep learning techniques to develop a perception stack for autonomous vehicles.

Additionally, a Tesla-inspired visualization was also developed using Blender to recreate a 3D scene from monocular images autnomously.

Implemented an algorithm to automate the estimation of camera intrinsics using a mathematical model proposed by Zhengyou Zhang.

Evaluated performance of state-of-the-art Gaussian Splatting method for real-time radiance field rendering on custom dataset featuring WPI's mascot 'Gompei'.

Implemented NeRF pipeline following this NeRF paper.

This project implements a Structure from Motion pipeline for obtaining camera poses from monocular images of a given scene and reconstructing the 3D scene by leveraging Epipolar geometry.

Implemented advanced Multi-State Constraint Kalman Filter (MSCKF) following a paper for robust sensor fusion.

Machine Learning

Aerial Robotics

- Engineered a perception stack using synthetic data and YOLOv8 for detecting and segmenting windows.
- Deployed the model on a DJI TelloEDU drone to autonomously navigate a complex environment in real-time.

- Implemented RRT path planner and cubic spline trajectory generator for quadrotor in 3D using Blender.
- Designed and tuned a PX4 based cascaded PID control for collision-free quadrotor navigation.
- Validated a robust collision detection system for real-time quadrotor planning.

Design of a sliding mode controller for altitude and attitude control of the Crazyflie 2.0 to enable the quadrotor to track desired trajectories and visit a set of desired waypoints.

Developed a dynamic 3D waypoint navigation algorithm and integrated AprilTag scanning for landing maneuvers.

Engineered a Cascaded PID controller based on the PX4 stack to ensure precise trajectory tracking.

Localization and Navigation

Performed robust quadrotor pose estimation in 3D space by using Quadrotor Dynamics and Extended Kalman Filter

Developed a vision-based observation model to get pose-estimates from AprilTags and improve prediction accuracy.

Implemented a Bayes filter for state estimation of a robot in a 2D space.

Implemented a Kalman Filter in Python for quadrotor localization in 3D space.

Implemented a Particle Filter (EKF) in Python for quadrotor localization in 3D space.

Robotics Projects

Position and Velocity control of SCARA robot using a PID Controller in a simulated environment in ROS2 and Gazebo.