Projects

Dexterous Robotics: Playing the Drums Using Deep RL

Project Overview

For my final project, I am working in a group to build on top of an already exisiting paper on dexterous robotics.

Implementation Details

We will take their design and implementation for piano and expand that to drumming. We will also focus on trying to improve the existing training methods noted in their paper to evaluate if we can achieve better results compared to the initial implementation noted in the paper.

Physical AI-Driven Action Sequence Analysis and MDP Modeling Using NVIDIA Cosmos Reason and Meta SAM2

code report

Project Overview

This project focuses on learning from video demonstration.

Implementation Details

The implementation uses Meta's SAM2 to correctly identify and segment the objects of interest to provide the neccessary contextual data that is later fed to NVIDIA's Cosmos Reason 8B model which will feed into the MDP (Markov Decision Process) design.

Robot Learning

code report

Project Overview

This project focuses on learning from demonstrated trajectories.

Implementation Details

The implementation uses DMPs (Dynamic Movement Primatives), DTW (Dynamic Time Warping), GMM/GMR (Gaussian Mixture Model/Regression) and methods for adaption to new trajectories and obstacle avoidance.

Autonomous Robotic Air Hockey

code video demonstration motor demonstration download project report

Project Overview

This project was done with a partner for our Advance Mechatronics I course.

Implementation Details

The implementation uses a gantry arm using NEMA-17 stepper motors controlled by an Arduino microcontroller. The system integrates sensors for puck detection and employs control algorithms to autonomously play air hockey against a human opponent. Computer Vision was handled through a ZED II camera, that used hue values to detect the puck and track its movement across the table. All of the computer vision and path planning computational load was handled by a Jetson Orin Nano.

Improving Image Classification Accuracy

code report

Project Overview

This project focuses on improving image classification accuracy on the CIFAR-10 dataset using various techniques.

Implementation Details

We take the original CIFAR-10 dataset and training code from the PyTorch documentation and from there see what techniques can be applied to improve accuracy.

Reinforcement Learning using Isaac Lab and Sim

code video

Project Overview

This project using NVIDIA Isaac Lab and Isaac Sim creates a reinforcement learning environment where robotic manipulators learn complex reaching tasks. The system trains robots to achieve precise pose and orientation targets through trial and learning. Using reinforcement learning we teach the manipulator to reach an end goal pose matching the end goal orientation.

A* Path Planning for a Supercharger network

code report

Project Overview

Our goal was to implement a search algorithm to find the minimum time path through a network of supercharging stations.

A* Path Planning

code video

A demonstration of A* path planning algorithm that dynamically renders its shortest path avoiding obstacles.

Project Overview

An interactive implementation of the A* pathfinding algorithm that visualizes the search process in real-time. The system demonstrates optimal path calculation while dynamically avoiding obstacles in a grid-based environment.

Path Planning Simulation

code

Particle Fluid Simulation that demonstrates path planning in dynamic environments.

Project Overview

This project explores path planning algorithms in dynamic environments that can be agitated through user interaction.

Implementation Details

Uses both SPH and Euler fluid simulation to create a dynamic environment that can be agitated through user interaction. This simulation uses both methods to create the particle behavior that then assigns two random particles to find the most optimal path between them while the environment changes.

Object Detection for VR

code Publication

A pipeline that gathers images from google maps API and converts them into VR friendly formats that runs an objection detection model that I labled for flammable objects within the picture.

Project Overview

An innovative VR application that combines Google Maps imagery with YOLOv8 object detection to identify flammable objects in real-world environments. This project creates immersive safety assessment tools for virtual reality platforms.

Particle Fluid Simulation

code video

Simulation of fluid dynamics using particles written in C++.

Project Overview

A high-performance particle-based fluid dynamics simulation built from scratch in C++. The system accurately models fluid behavior using computational physics principles and provides real-time visualization of complex fluid interactions.

RobotBallet GoogleDeepmind

code Google Deepmind's whitepaper

My attempt at replicating the research presented by GoogleDeepmind in MuJoCo. This is currently still on going.

Project Overview

This project focuses on replicating the groundbreaking robotics research from GoogleDeepmind, specifically their work on creating ballet-like movements in robotic systems using advanced motion planning and control algorithms.

Implementation Details

The implementation uses MuJoCo's advanced physics engine to create realistic robotic movements that mirror the grace and precision of ballet. The system incorporates complex trajectory planning algorithms to ensure smooth, flowing motions while maintaining physical constraints and stability.