Featured Projects
Robotics projects in perception, control, and autonomy
MonoSense: Monocular Perception Pipeline
FEATUREDApril 2026
Full autonomous-driving perception pipeline combining a custom YOLO detector, DepthAnythingV2, FCOS3D 3D pose estimation, UFLDv2 lane detection, and RANSAC ego-motion.
Key Achievements
- Custom YOLO reached 0.564 mAP50 on BDD100K and 0.859 mAP50 on LISA traffic signs
- Rendered 27,000 synthetic frames across 13 urban scenes on a SLURM-managed A30 GPU cluster
- A 1D Kalman filter on the vanishing point cut lane projection variance 60% on curved roads
- Integrated DepthAnythingV2 for metric depth estimation and FCOS3D for 3D vehicle pose estimation
March to April 2026
Deep Visual Inertial Odometry
Deep learning VIO stack for UAVs with 6 fusion approaches on custom Blender rendered trajectories.
- MSCKF reached 0.12m ATE RMSE over a 73m EuRoC trajectory, a 0.21% drift rate
- Implemented 6 VIO approaches: vision only (DeepVO style), AirIMU+AirIO, PRGFlow, PRGFlow+yaw, EKF loose fusion, cross attention tight fusion
- Built vision only odometry using Siamese MobileNetV2 + correlation + LSTM for relative pose estimation
April 2026
Imitation Learning for Robotic Stacking
Behavioral Cloning vs Diffusion Policy comparison for UR5e robotic stacking with keyboard teleoperation.
- BC-Transformer reached 100% rollout success by epoch 950
- Diffusion Policy reached 80% success by epoch 350, with non-zero success from as few as 5 demonstrations
- Implemented BC Transformer and Diffusion Policy for UR5e robotic stacking task in robosuite simulator
March 2026
Deep Reinforcement Learning for Robotic Picking
Implementation of deep RL algorithms (REINFORCE, Actor Critic, A3C) for robotic manipulation tasks.
- Benchmarked REINFORCE, A2C, and A3C on LunarLander-v2; A2C reached mean reward 201.7 vs. 175.3 for REINFORCE, with a 29% cut in reward variance versus REINFORCE
- Deployed A3C with a CNN actor-critic on PyBullet Kuka, reaching 31% grasp success over 100 episodes
February to March 2026
Structure from Motion and Neural Radiance Fields
Implemented classical Structure from Motion pipeline from scratch and trained Neural Radiance Fields for photorealistic 3D scene reconstruction.
- Implemented complete SfM pipeline: RANSAC feature matching, 8 point fundamental matrix, essential matrix decomposition, cheirality based pose disambiguation, linear/nonlinear triangulation
- Built PnP solver with RANSAC for camera registration from 2D to 3D correspondences
- Reduced reprojection error by 36% (from 33.2 to 21.2 pixels) using sparse bundle adjustment
November to December 2025
Quadrotor Trajectory Control
Implemented PD and LQR controllers in PyBullet simulation, then validated on real Crazyflie 2.0 hardware. Performed system identification reducing sim to real RMSE from 4.8mm (simulation) to 0.8mm (real hardware), an 83% improvement. LQR controller maintained position error below 2cm on hardware. Also implemented polynomial trajectory generation for smooth 3D paths.