Priyanshu Mishra ([email protected])
New York University, Abu Dhabi
New York University
This research project explores the integration of advanced robotics and machine learning techniques to develop an autonomous system capable of navigating complex environments and performing object manipulation tasks. Utilizing a library setting as a case study, the project implements a novel approach combining reinforcement learning for navigation with computer vision-based manipulation.
The navigation system employs the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, an advanced reinforcement learning technique that enables continuous control in high-dimensional action spaces. This allows the robot to learn optimal pathfinding strategies in dynamic environments. Simultaneous Localization and Mapping (SLAM) techniques are integrated to enhance the robot's spatial awareness and adaptability.
For object manipulation, the system incorporates a 6-degree-of-freedom robotic arm controlled through inverse kinematics. Computer vision algorithms process RGB-D camera data to facilitate object detection and grasping planning.
A key innovation is the development of a high-fidelity simulation environment based on LiDAR-scanned data from real library spaces. This environment, implemented in Gazebo, allows for rapid iteration and learning, addressing the challenge of sim-to-real transfer in robotics.
The project demonstrates the feasibility of deploying autonomous systems in structured public environments, showcasing promising results in navigation accuracy and manipulation precision. This research contributes to the growing field of autonomous systems and offers insights into the technical challenges and ethical considerations of implementing such technologies in public service roles.
Below you will find the PDF of my Capstone Research Project. Feel free to view or download it for a comprehensive look at the paper.
CapstoneDesignDoc - pm3013 (1).pdf
This capstone project explores the implementation of autonomous navigation systems using Reinforcement Learning from Human Feedback (RLHF), with potential applications extending beyond the initial library use case. The project aims to develop a robust, adaptable robotic system capable of navigating complex environments and performing manipulation tasks. Since the research project was completed in 2023, I’ve had the opportunity to look over it a few times and try to summarize the key technical components of the paper for anyone reading this in the future.