Hello, I'm Kevin

Led the US deployment of the EVE Autonomy platform, developed with Yamaha and Tier IV, focusing on ROS2-based SLAM and localization, integrating 3 Livox Horizons, a Velodyne VLP-16, and RTK for advanced motion planning. Engineered a detailed HD map of the Pennovation site with lanelets using Tier IV tools, implemented and stress-tested multiple behavior planners for robust performance, and advanced navigation accuracy with Visual Inertial Odometry integration.

Note: For an optimal experience, use browsers on laptop/PC

View on GitHub

Led a 3-day robotics hackathon as the sole U.S. team, building an autonomous robot using ROS2, Gazebo, and Olive Robotics’ hardware. Developed a custom localization stack, integrating odometry and AprilTags for precise navigation. Achieved 85% accuracy in autonomously locating and stacking bento boxes.

Note: For an optimal experience, use browsers on laptop/PC

View on GitHub

Presenting an innovative autonomous 4-wheeled mobile robot that seamlessly integrates mechanical design, electronic engineering, and sophisticated embedded software. This robot is primarily capabable of precise wall following, beacon tracking, and pushing a model police car (a heavy target placed randomly on the track). The fundamental distinguishing component used are the mecanum wheels which provides omnidirectional motion to the robot!

The brains of the robot is the ESP32 S2 microcontroller, which works as the access point facilitating real-time data processing crucial for managing interactions and executing tasks. The robot employs three crucial circuits:

• frequency detection circuit which uses LTR-4208 phototransistors to detect IR LEDs (beacon), signals processed using ATMega32u4
• HTC Vive circuit which uses PD-70 photodiode to read and transmit robot's position, used for robot localisation PID algorithm
• ultrasonic sensors and motor drivers (TB6612FNG) circuits used to drive the robot

The software architecture plays a pivotal role in the robot's functionality, incorporating a real-time operating system (RTOS) to efficiently handle concurrent tasks, ensuring seamless and responsive operation. The system employs interrupt handling for immediate response to sensor inputs, I2C communication for streamlined data transfer between sensors and the microcontroller, and UDP, ESP-NOW protocols for robust network communication.

Note: For an optimal experience, use browsers on laptop/PC

View on GitHub

This project tasks a 7-DOF Franka Emika Panda arm to pick static blocks and dynamic blocks and place them in a stack at the target platform. The key task here is strategizing the best motion planning algorithm. As for the competition stacking the blocks in a particular way is critical which requires adept handling of both static and dynamic blocks. Accurate detection of the blocks using the camera (at end-effector) and manipulating the end-effector were the pivotal tasks. The final algorithm developed here seamlessly integrates Geometric Inverse Kinematics and a bi-directional Rapidly-exploring Random Tree (RRT), ensuring precise and efficient control over the robotic manipulator.

A notable milestone is the successful development of an analytical Inverse Kinematics (IK) solution, significantly enhancing the speed and efficiency of the robotic manipulator compared to traditional IK solutions by bypassing iterative steps- an imperative factor for real-time applications. The project also serves as a testament to practical problem-solving skills and adaptability, especially in a competitive environment.

Note: For an optimal experience, use browsers on laptop/PC

View on GitHub

The Neural Radiance Field (NeRF) uses a multilayer perceptron (MLP) to learn a compact implicit neural representation of a static 3D scene from multiple input views. It maps 3D coordinates and view directions to color and density values. By neural volumetric rendering of many sampled points along camera rays, it can reconstruct input views and also render novel views of the scene. This project involves fitting 2D and 3D scenes using neural networks.

2D Image Fitting:

• An MLP is trained to map 2D pixel coordinates to RGB values to fit a given image. Positional encoding is used to map inputs to a higher dimension. The MLP has linear layers with ReLU activations except the last sigmoid layer.
• The model is trained to reconstruct the image. Positional encoding helps achieve higher reconstruction quality.

3D Scene Fitting:

• An MLP takes 3D coordinates and view directions as input and predicts color and density values to represent a 3D scene.
• Rays are cast through pixels of input view images. Points sampled along each ray are fed to the MLP.
• Volumetric rendering of the MLP outputs allows reconstructing input views and rendering of novel views.
• The model is trained end-to-end on input views to get a compact scene representation for view synthesis.

Note: For an optimal experience, use browsers on laptop/PC

View on GitHub

This project immerses in computer vision, focusing on essential matrix estimation, relative pose disambiguation, triangulation, and reprojection error analysis for multiview 3D reconstruction.

Estimating the essential matrix (E) involves capturing epipolar constraints between matching points. Refined through least squares and RANSAC, E unlocks the relative camera pose. Triangulation reconstructs 3D structures, resolving pose ambiguity with SVD. Evaluation based on point positioning enables recovery of relative rotation (R) and translation (T). Reprojection errors provide critical insights into the accuracy of reconstructed 3D points.

Note: For an optimal experience, use browsers on laptop/PC

View on GitHub

This project delves into dynamic scene analysis by computing optical flow between images in a sequence, utilizing spatiotemporal image derivatives and a local constant flow assumption. The computation involves Gaussian derivative filters to obtain image derivatives along x, y, and t, subsequently solving a local 25x2 linear system to derive the 2D flow vector (u,v). Visualizing the flow as a vector field facilitates the computation of the episode of the dominant motion using the epipolar constraint.

With the optical flow and epipole known, the motion is further dissected into rotation and translation components. By rearranging the projection equation, depth values are calculated at each pixel, creating a depth map visualization through the structure from motion.

The key learning outcomes include hands-on experience with optical flow computation, utilizing epipolar geometry for scene structure reasoning, and recovering depth from camera motion. In essence, this homework provides practical skills applicable to video analysis, 3D reconstruction, robotics, and autonomous vehicles.

Note: For an optimal experience, use browsers on laptop/PC

View on GitHub

This project focuses on building an augmented reality application using AprilTags for camera pose estimation, reinforcing concepts of homographies and coplanar constraints. The key task is to leverage the AprilTag fiducials to estimate the camera pose in different ways.

First, a world coordinate system centered on the tag is established and pose is estimated using the Perspective-N-Point (PnP) algorithm. Next, the Perspective-Three Point (P3P) algorithm combined with Procrustes analysis provides an alternative pose estimation approach. Converting 2D pixel picks to 3D world coordinates also highlights the interplay between camera calibration and geometry.

By putting together these building blocks, geometric CV principles like camera projection models, pose estimation, and augmenting real scenes with virtual objects was implemented, giving a firm understanding of concepts needed for mixed reality applications.

Note: For an optimal experience, use browsers on laptop/PC

View on GitHub

Frames of a soccer video is augmented by projecting a logo onto the goal area in each frame. The homography between points on the goal and logo are estimated using the Direct Linear Transform (DLT) algorithm, and to warp the goal points onto the logo plane. The warped correspondence is then used to copy logo pixels into the video frame.

Key concepts include interpreting homographies as transformations between 3D world and 2D image spaces and using four point matches to set up and solve equations for the inter-image homography matrix. Inverse warping avoids overlapping projections and holes.

Note: For an optimal experience, use browsers on laptop/PC

View on GitHub

Developed effective obstacle avoidances algorithm for an autonomous 6 - DOF quadcopterdrone in an unknown environment using the concept of Nagumo's invariance theorem, viz., using safety velocitycones (Bouligand tangent cones). Then tested the newly developed deadlocks - free hybrid approaches on a fully customized small (2 - 25kg weight) dronemodel- Holybro X500 V2

Note: For an optimal experience, use browsers on laptop/PC

View report

This is my BTech thesis project under Dr. Satyanarayan Panigrahi. I am developing control schemes using L1 adaptive control to project obstacle avoidance and guidance of drones to actual(fixed-wing) UAVs. Further, a robust adaptive control system for the autonomous landing of quadrotors will be developed.

Note: For an optimal experience, use browsers on laptop/PC

View report

Developed a working model of the steering mechanism used in SAE BAJA's buggy as per given dimensional regulations using SolidWorks. This model was crucial in determining the material needed and for further CAE analysis in Ansys.

Further models of roll cages were developed by adding some members and trusses in Solidworks. ANSYS was used to create a finite element formulation of the problem for both static structural and dynamic analysis. Then maximum stress and deformation of frontal, side, rearand rollover of different models were analysed

Note: For an optimal experience, use browsers on laptop/PC

Developed a full - stack application with features similar to the Slack app built using React, Redux, and Firebase. The app has basic authentication,proper chatting channels, image upload in chat features, emojis, and much more.

It is a live application hosted using Firebase. Sample user entry- Email: 123@gmail.com, password: 123456 , else register for a new user

Note: For an optimal experience, use browsers on laptop/PC

Visit site or View Source

During the lockdown period I was up for a challenge. So I decided to wish my mother in a more modern way by creating this webpage for her. Feel free to use this page to surprise your loved ones in a new way!!

Note: For an optimal experience, use browsers on laptop/PC

Visit site or View Source