Time-Optimal Re-planning of Quadrotor Trajectories

Abstract

With the rise of quadrotor drones in recent years, the research and development of time-optimal trajectory planners are pushing the boundaries. They now not only exploit the full dynamics of the drone to generate aggressive trajectories but also have runtimes that allow them to generate plans in near real-time.

This work extends current state-of-the-art time-optimal quadrotor trajectory planners to allow for on-the-fly trajectory re-planning. Given new waypoints and a previous trajectory, the planner is able to generate an updated trajectory while maintaining time optimality.

Because the planner leverages a learned sequence to sequence neural network model, it is able to generate trajectories magnitudes faster than optimization based approaches. This work then takes it one step further and optimizes the planner using a compiled real time inference library (NVIDIA TensorRT). The optimized planner is demonstrated to provide a 14.84 times increase in throughput and over 95% reduction in latency. The increase in throughput can be translated to better efficiency, and the reduction in latency is critical for trajectory re-planning where the drone is flying an active trajectory. Both improvements push it one step closer to running the planner onboard drones themselves.

Although most experiments were conducted on desktop class hardware, mobile chips like the NVIDIA Jetson AGX Orin were also tested to mimic the class of hardware that could be flown onboard drones.