active
Drone AI
Autonomy stack for a small quadrotor — state estimation, motion planning, and a learned control policy built from scratch.
A from-scratch autonomy stack for a quadrotor platform. The system handles state estimation from IMU and optical flow data, trajectory planning through cluttered environments, and low-level attitude control.
State estimation fuses inertial measurements with visual odometry in an extended Kalman filter. The planner uses a sampling-based approach to generate feasible trajectories, which a tracking controller then follows.
The longer-term aim is to replace the hand-tuned controller with a neural policy trained on simulated flight data. For now the classical pipeline flies reliably and provides the baseline that any learned controller must beat.