To achieve autonomous operations, a UAV must be able to plan motions and trajectories. These should ideally exploit the vehicle’s full dynamic capabilities, be fast to compute even on limited computational hardware, and be also useful outside of perfectly instrumented laboratory environments.
We are interested in reliable and safe operation in unknown, unstructured outdoor environments. We have developed a memoryless local planner, that plans motions in the output of a depth camera. Each camera frame represents a map of the vehicle’s immediate environment; we generate a new motion plan with every camera frame. The result is a receding horizon planner that naturally accounts for changes in the environment.
We are also interested in the ability to avoid collisions with dynamic obstacles. The below video shows some of our work, allowing a drone to duck out of the path of a thrown obstacle (though we here rely on an external motion capture system for estimating the state of the vehicle, and dynamic obstacle).