MS Robotics Engineering @ WPI | Perception • Robot Learning
📧 raghavnallaperumal753@gmail.com 🌐 Portfolio | LinkedIn
I build intelligent robotic systems that perceive, learn, and act in the real world. My work spans vision-based perception (MonoSense, VIO, NeRF), learning-based manipulation (deep RL, imitation learning), and real hardware deployment (Franka Panda, UR5e, Crazyflie).
Graduate Research Assistant, Aerial-robot Controls and Perception Lab (ACP Lab), WPI, advised by Dr. Guanrui Li. Building a monocular depth prediction network that generates information for reinforcement learning exploration policy, so a quadrotor can navigate unknown, maze-like environments without depending on information that isn't available once it's actually deployed.
🚗 MonoSense – Autonomous driving perception pipeline: custom YOLO (0.564 mAP50 on BDD100K, 0.859 mAP50 on LISA traffic signs), DepthAnythingV2, FCOS3D 3D pose, UFLDv2 lane detection. 27,000 synthetic frames rendered across 13 scenes on a SLURM A30 cluster; a 1D Kalman filter cut lane projection variance 60% on curved roads.
🛸 Deep VIO – 6 visual-inertial odometry approaches compared for UAVs, including MSCKF, a 15-state EKF, a DeepVO-style LSTM, and a cross-modal transformer. MSCKF reached 0.12m ATE RMSE over a 73m EuRoC trajectory (0.21% drift).
🤖 Imitation Learning – BC-Transformer vs Diffusion Policy for UR5e contact-rich block stacking, trained on 158 human teleoperation demos. BC-Transformer hit 100% rollout success by epoch 950; Diffusion Policy reached 80% by epoch 350 and stayed non-zero with as few as 5 demos.
🎮 Deep RL for Picking – REINFORCE, A2C, and A3C built from scratch. A2C cut reward variance 29% versus REINFORCE (201.7 vs 175.3 mean reward); A3C reached 31% grasp success on a PyBullet Kuka arm.
📐 SfM + NeRF – 3D reconstruction from scratch: bundle adjustment cut reprojection error 36%, NeRF training reached 27.3 dB PSNR / 0.90 SSIM.
🚁 Quadrotor Control – PD and LQR control on a Crazyflie 2.0. System identification cut sim-to-real RMSE 83% (4.8mm to 0.8mm); LQR held hardware position error below 2cm.
Languages: Python, C++, CUDA, MATLAB Robotics: ROS 2 (Humble), MoveIt 2, Nav2, Gazebo, PyBullet, robosuite, robomimic, Simulink Machine Learning: PyTorch, Deep RL (A2C, A3C), Imitation Learning, Diffusion Policy, World Models Perception: OpenCV, Kornia, YOLO, Depth Estimation, VIO, SLAM, Structure from Motion, NeRF, Camera Calibration Controls: PID, LQR, MPC, EKF, System Identification, Trajectory Optimization Hardware: Franka Emika Panda, UR5e, Crazyflie 2.0, Intel RealSense D435, Beckhoff TwinCAT PLC, NVIDIA Jetson Nano Tools: Docker, Git, SLURM/HPC (A30), Blender, HM3D, Habitat-Sim
📊 Seeking Full-Time Opportunities (Aug 2027) in perception engineering, computer vision, robotics software, robot learning, or autonomy.