Bimanual manipulation using Diffusion Policy 🔗

• Implemented and trained a diffusion policy for bottle uncorking task using a dual UR5e arm setup.
• Utilized ROS for data collection and PyTorch for model implementation.
• Achieved an average task completion rate of 74.7% across 30 rollouts.

Collision Avoidance for Dual UR5e Arms Setup 🔗

• Developed a collision avoidance module for the teleoperation system of dual UR5e robot arms.
• Utilized Vector-Accelerated Motion Planning to perform collision checks at 1000 Hz using data from the robot's proprioception.
• Reduced the number of emergency stops due to collisions by 95% for new users.

Scene Description for the Visually Impaired 🔗

• Designed a navigation system to assist visually impaired users.
• Utilized YOLO for object detection, LiDAR for depth input, and an LLM to provide coherent language instructions.
• Achieved an 81.4% net preference score over baseline VLM.

Self Driving Vehicle Lateral and Longitudinal Control 🔗

• Implemented self-driving vehicle control in CARLA simulator to control a vehicle given waypoints.
• Implemented PID controls for longitudinal control; and Pure Pursuit & Stanley controls for lateral control.
• Developed sensor fusion with ES-EKF for state estimation, integrating IMU, GNSS, and LiDAR data.

Grasping of Dynamic Object under Dynamic Scenarios 🔗

• Designed a grasping pipeline for dynamic objects marked with ArUco tags, capable of handling occlusions and obstacles.
• Implemented the system on a dual UR5e arm setup, utilizing OpenCV for ArUco detection and VAMP for collision detection.
• Achieved smooth motion for grasping by incorporating a confidence-based movement metric.

Human Pose Imitation on Baxter Robot 🔗

• Developed a system to replicate real-time human actions captured via a camera on a simulated Baxter robot.
• Utilized MediaPipe for 3D keypoint extraction and designed an algorithm to map keypoints onto the robot.
• Built a live demo website, demonstrating consistent performance and adaptability to various human shapes and sizes.

Disassembly of Articulated Objects Using Dual Arm Robot 🔗

• Developed an object disassembly system for a dual UR5e arm setup in PyBullet with two teammates.
• Created a segmentation-guided grasping module for generating 6-DOF grasp poses using SAM and ContactGraspNet.
• Designed a deep learning model to classify disassembly operations from videos using features extracted with MediaPipe.

Human Hand Tracking and Pose Estimation 🔗

• Led a team of 4 graduate students to develop a hand tracking and pose estimation model with ResNet50 as backbone.
• Trained the model on the HaGRID dataset using a YOLO like loss function, achieving a mAP@50 of 0.68 on the test data.
• Created an interactive virtual board using OpenCV, enabling users to draw using hand gestures @8 fps.

Movie Recommendation System Using Graph Convolution Network 🔗

• Implemented a Graph Convolution Network (LightGCN) using PyTorch Geometric to develop a movie recommender system.
• Transformed the dataset into sparse tensors and applied Bayesian Personalized Ranking loss function for model optimization.
• Achieved a recall@20 of 0.12 on the MovieLens dataset during testing.

AI agent for Ultimate Tic-Tac-Toe 🔗

• Developed a human playable Ultimate Tic-Tac-Toe game in Python3, leveraging object-oriented programming principles.
• Implemented an AI agent capable of searching a state space of size 500k+ in <1sec using MiniMax algorithm with pruning.
• Designed custom heuristics for the AI, achieving a 70% win rate against online available AI opponents.

Web Crawler for Notifying Driving License Test Availability 🔗

• Developed a web crawler to monitor the Minnesota DVS website for test availability every 5 minutes.
• Employed Selenium to automate the input of user details and navigation of the webpage to check availability.
• Incorporated email notifications into the script to alert me when an appointment slot becomes available.

Interview with a Prominent Indian Faculty Providing Tips for BARC Interview Preparation

Presentation on Diffusion Policy for Robot Manipulation