Robots in Service of the Enviroment (RSE)

• Migrated the software to ROS2, achieving a notable 40% reduction in cpu usage. Implemented a Linux to ARM cross-compiling pipeline which reduced build time by 20%.
• Trained YOLOv7 and Deep SORT for Lionfish detection and tracking in marine coral regions achieving an 80% detection accuracy with 0.89 mAp at 30 fps tracking inference on the robot.
• Developed a GUI using PyQt5, experimented with threading and asyncio to address processing bottlenecks.
• Tools: Python, C++, ROS, ROS2, Pytorch, YOLOv7, DeepSORT, RaspberryPi-4b, gstreamer, PyQt5

Vision, Intelligence, and System Laboratory(VIS Lab)

• Proposed a custom network utilizing both PENet and DPT for dense depth prediction from sparse depth inputs of Lidar, and tested on KITTI Dataset to achieve a 5% decrease in local RMSE.
• Experimented with Sparse Convolution dilation rates to reduce network parameter size for online applications
• Tools: Python, Pytorch, DPT

Human-Inspired Robotics Lab (HIRO Lab) WPI,

• Implemented D* lite algorithm for navigation and SRRT for manipulation and deployed on LocoBot, a differential drive robot with Mobile PX100 manipulator, a 6dof arm using Moveit and packaged in a docker container.
• Performed an incremental experiment with Deep Reinforcement Learning for autonomous navigation, starting with Q-learning followed by actor-critic model using openai-ros package. Trained A2C model using TensorFlow in Gazebo for end-to-end planning with vision(Realsense d435) and odometry inputs.
• Mentored a team of 2 members to first simulate the behavior and then replicated the results on hardware.
• Tools: Python, C++, ROS, Open3d, LocoBot, D* lite, RL, Rviz, Gazebo

Ubicept

• Highlighted superior performance of proprietary SPAD sensor imaging compared to CMOS sensors in applications of Visual SLAM, achieving a 15% better trajectory tracking accuracy and 40% regaining tracking recovery rate.
• Developed a benchmarking setup using Turtlebot3 and Jetson to pipe camera inputs via ROS bag for offline processing.
• Created a calibration setup for monocular and mono-inertial setup, also simulated SPAD in Unity Environment.
• Tools: python, ROS, Linux, Nvidia Jetson, OrbSLAM3, DSO, 3D Printing

Languages and Databases

Python

Shell Scripting

Libraries

NumPy

Pandas

OpenCV

scikit-learn

matplotlib

Frameworks

TensorFlow

PyTorch

Other

Git