Teleoperated and trained a ARX humanoid robot to perform three different tasks for with bimanual manipulation: Fold a table cloth, open a container, and pick and place an object. Policy was trained using ACT, and competition was organized by Google Deepmind, RAI Institute, and other leading partners.

Designed, constructed, and trained a locomotion policy for a six-legged robot to compare learned gaits to traditional ones. Programmed a tripod gait with inverse kinematics and bezier curves, while the reinforcement learning used PPO (Proximal Policy Optimization) and the Genesis Simulation platform.

A 7-DoF Franka robot arm that adjusts its z-height using AprilTags, detects faces with an external camera, and generates portraits via line art. My core contribution involved leading the perception subsystem, utilizing ROS2, OpenCV, and Python to convert source images into detailed line drawings.

Trained a Unitree Go2 robot dog in Genesis to perform diverse locomotion tasks, including walking, running, jumping, strafing, and stair climbing. Each specialized policy was developed using Proximal Policy Optimization (PPO) with task-specific reward functions.

Created GPT-2's transformer from scratch adapting the source code from Andrej Karpathy's tutorials. Trained on 2 A6000 RTX GPUs with a total training time of 4 days. Evaluated results on hellaswag data retrieving an accuracy that beat the original model.

Integrated a VLM (Vision-Lanugage Model) with SAM (Segment-Anything Model) to detect and mask arbitrary objects in a scene. This mask is overlayed onto the point cloud provided by a realsense RGBD camera and filters out grasping poses calculated from GraspNet.

Autonomous navigation system for an RC car using imitation learning built on a custom CNN (convolutional neural network).

Interactive path planning tool for navigating occupancy grids created by SLAM algorithms. Implemented the A* algorithm in C++ and integrated it with ROS2, allowing users to define start and goal points and visualize optimal paths in a simulated environment.

Implemented the Jancey K-Means algorithm in C++, comparing this lesser-known variant of K-Means to other popular widely used algorithms for clustering applications.

Custom CNN (Convolutional Neural Network) that predicts a live drawing from a user over 250 unique classes.

Simulated a mobile Kuka robot in Coppelia Sim performing a manipulation task with a 4 DoF arm attached. Utilized trajectroy generation and PID control for smooth planning.

Simulated a jack in the box using Langragian dynamics and impact equations with Python and sympy.