Khuzema Habib

Master of Engineering in Robotics

University of Maryland, College Park | 2024 - 2026

Bachelors of Technology in Mechanical Engineering

Manipal Institute of Technology | 2019 - 2023

Autonomous Mobile Manipulator

Built and developed a fully autonomous robot capable of computer vision-based object localization and servo-controlled manipulation.

• Built and Assembled a Mobile Robot from Scratch with an integrated gripper for mobile manipulation tasks.
• Raspberry Pi 4B platform with a camera and ultrasonic sensor, along with an Arduino nano for IMU data collection.
• Integrated sensor-fused navigation with wheel odometry and IMU data.
• Pick-and-place tasks with subsystems integration.
• Successfully completed the final challenge of the competition, where the robot needed to autonomously find RGB blocks and place them in a target zone.
• Added a custom designed LED ring indicator to the robot for navigation state visualization.

Autonomous Vehicle Behaviour Planning In CARLA

Worked on a behavior planner for autonomous vehicles in CARLA with Behaviour Trees for decision making.

• Behavior Planning: Uses Behavior Trees to decide high-level maneuvers (LANE_KEEP, FOLLOW, LANE_CHANGE, STOP_FOR_LIGHT, RECOVERY).
• Trajectory Planning: Generates candidate trajectories in Frenet coordinates and selects the optimal path based on a weighted cost function.
• Vehicle Control: Implemented PID controller for longitudinal speed and Stanley controller for high-fidelity lateral steering.
• State Machines: Integrated tracking for in-progress lane changes and a robust recovery mode for off-road or stuck scenarios.

Stanford Doggo Inspired Quadruped

Ongoing

Building an agile and durable quadruped robot inspired by Stanford Doggo to use as a testbed for future projects

• Fabrication (3D Printing and CNC machining) and Assembly
• Wiring and Odrive setup with custom firmware, Teensy 3.5 Microcontroller for control with Xbee for Wireless communication
• Chassis reinforcement with the use of Carbon Fiber Rods over Sheet Metal
• Testing and fine-tuning the build to ensure proper functionality - Need to fix some issues with wea k er PLA components snapping under load when walking
• Future integration with ROS2 for Autonous Navigation and easier integration with other sub s ystems
• Future improvements to mechanical design to add additional degree of freedom to the legs

A Star Path Planner for Turtlebot3

Developed an A* path planner for a Turtlebot3 robot to navigate through a maze-like environment. Validated on Gazebo Simulation as well as real-world testing.

ROS 2 Python Gazebo Turtlebot3 A* Path Planner
• Developed a 2D map environment using half-planes and semi-algebraic equations with configurable robot clearances.
• Integrated a trajectory controller using proportional control to translate planned waypoints into linear and angular velocities.
• Validated path execution across multiple simulation platforms, including Gazebo, the Falcon Simulator, and Matplotlib.
• Processed real-time odometry data (/odom) to handle pose estimation and heading angles via quaternion transformation.
• Optimized movement by incorporating differential drive constraints, allowing user-defined RPM inputs for wheel velocities.

Pursuit RRT* Path Planner for Turtlebot3

Developed an RRT* path planner for a Turtlebot3 robot to navigate through a maze-like environment to catch a moving target.

• Developed a 2D RRT* path planner for a maze-like environment to catch a moving target.
• The planner implements a pursuit strategy to catch a randomly moving target by using RRT* to plan a path to the target's current position.
• Extended this to a Gazebo simulation with Turtlebot3 robots, where the pursuer robot uses RRT* to plan a path to catch the target robot.
• The pursuer robot implements extracts the current pose of the evader robot and plans a path, and is considered caught when the robot is in view and a certain distance away using a depth camera.
• The Evader robot implements proportional control to follow the waypoints, subscribing to /odom to extract pose and heading angles from the quaternion, then publishing linear and angular velocities to /cmd_vel

Multi-Agent Reinforcement Learning for Drone Swarm Control

Custom MAPPO and IPPO Implementation for Advanced Coordination Tasks in gym-pybullet-drones environment.

• Developed custom Multi-Agent Proximal Policy Optimization (MAPPO) and IPPO algorithms from scratch using PyTorch for high-efficiency swarm control.
• Extended the gym-pybullet-drones environment to include complex multi-agent scenarios: hovering, spiral formation, and leader-follower navigation.
• Implemented a centralized critic in MAPPO to resolve global credit assignment, significantly reducing return variance during training.
• Achieved a 3x performance improvement over Stable-Baselines3 (SB3) in coupled tasks, reducing Spiral RMSE to 0.25m.
• Optimized safety metrics, reaching a low collision rate of 2.1% while maintaining maximum episode stability of 240 steps.
• Eliminated the overhead of standard RL libraries (Rllib/MARLlib) by building a streamlined, flexible training pipeline.

LQR Controller for Double Inverted Pendulum Cart Problem

Implemented an optimal control strategy for a complex nonlinear system involving a cart and two coupled pendulums.

• Derived the nonlinear equations of motion for a cart with two pendulums using Eul e r-Lagrange equations.
• Linearized the complex nonlinear system about a specific equilibrium point to create a state-space representation.
• Developed an optimal LQR controller by penalizing state error and actuator effort thr o ugh tuned Q and R matrices.
• Certified the closed-loop stability of the system using Lyapunov's indirect method and eigenvalue analysis.
• Designed Luenberger observers for observable output vectors to estimate system states fro m limited measurements.
• Implemented a Linear Quadratic Gaussian (LQG) controller to maintain optimal performance in the presence of process and measurement noise.

Visual Servoing with MPC

Designed MPC visual servoing controller for Turtlebot3 navigation using RealSense and Aruco markers.

• Model Predictive Control: Integrated a visual servo-based MPC framework to steer nonholonomic wheeled robots.
• Polar Kinematics: Transformed robot kinematics into a polar coordinate frame to ensure smoother steering and avoid Cartesian singularities.
• Real-time Optimization: Utilized CVXPY and OSQP Solver to solve Quadratic Programming (QP) optimization problems efficiently.
• Visual Servoing: Used a depth camera for position-based visual servoing and real-time depth feedback.
• Constraint Enforcement: Enforced operational constraints such as actuator saturation and velocity limits directly within the control design (Q and R matrices).
• Gazebo Validation: Validated the framework in simulation, demonstrating superior precision and speed over traditional PD control.