I am a final-year Robotics & Mechatronics Engineering student at the University of Dhaka. My research builds adaptive neural controllers that let robots keep moving under uncertain, deformable contact; currently, learning policies that recover a Mars rover's mobility when its wheels become entrapped in granular terrain, where classical control fails under slip.
Methodologically I focus on reward shaping with safety-gated constraints, for instance, yaw-projected rewards with entrapment-gated slip penalties, and on Sim2Real transfer through domain randomization and curriculum learning. More broadly, I am drawn to safe and preference-driven reinforcement learning, learning-based control, and neuro-robotics for movement: robots that move robustly in the physical world and learn from interaction in safe, sample-efficient, human-aligned ways.
Wheel entrapment in loose regolith has crippled real planetary missions (e.g., NASA's Spirit rover). We cast recovery as a reinforcement-learning control problem in deformable granular dynamics, learning a maneuver-agnostic policy that exploits terrain feedback to regain mobility without a scripted escape recipe, bridging RL, locomotion, and safe control under uncertainty.
A bare-metal flight controller on an STM32F103, closing six control loops at 250 Hz with no RTOS and no flight stack. Five sensors fused into an attitude and position estimate, feeding a rate-PID cascade. Built in 2023 and flown and upgraded across my undergraduate years.
A full kinematics and dynamics stack written from first principles in NumPy, validated against official URDFs for both a UR5e and a 7-DOF KUKA iiwa. Cross-checking exposed a DH sign error worth 1.998 m; corrected, model and URDF agree to 1.5e-7.
A software-in-the-loop maritime-patrol drone over Chattogram port: PX4 SITL and Gazebo render the flight while a ROS 2 planner routes exclusion-aware waypoints that never enter a no-fly zone, driven from a browser ground station with a live chase-camera feed.
An APPO agent trained with Sample-Factory to survive in a first-person 3D environment, learning to navigate and collect health packs from raw pixel input.
A 31.8 × 54 m digital-twin warehouse benchmarking PPO, SAC, and TD3 for autonomous navigation on a Clearpath Jackal under realistic LiDAR and localization noise.
A vectorised Python flocking simulation with five steering rules, measured order parameters, and obstacle avoidance. Boids coloured by heading show alignment as colour convergence — a rainbow collapses to a single hue.