Autonomous Quadrotor Navigation, Perception & Control
Building a navigation and control framework for a quadrotor to fly through a sequence of racing gates, then extending it to onboard stereo-vision navigation through openings it has never seen, removing the need for an external motion-capture system.
Completed
- Learning-based control using Proximal Policy Optimization (PPO), with the policy generating control actions from gate-relative state.
- Navigation formulated on a kinematic quadrotor model, regulating motion relative to the target gate's position and orientation.
- Actuation and state-estimation delays modelled in simulation, with domain randomization for sim-to-real transfer.
- Trained controller deployed and flight-tested on a Crazyflie 2.0, using Qualisys motion capture and the Crazyflie EKF for state estimation.
In progress
- Onboard perception using drone-mounted stereo cameras, to replace external motion capture.
- A gate and opening detection pipeline from stereo depth, estimating position, orientation and geometry relative to the quadrotor.
- Extending the framework from fixed racing gates to arbitrary, previously unknown openings.
- Largest-rectangle fitting on detected openings to extract the largest feasible passage, fed back to the controller as opening-relative state.