The concept of an autonomous racing drone has been researched for several years. The DRL tested an AI-enabled racing drone that could beat human pilots. Obstacle-avoidance and spatial awareness is increasingly getting better. A team from the University of Zurich (UZH) has taken this to a new level. They demonstrated an autonomous drone that flew through a dense forest at 40kph without crashing. This marks a new milestone for autonomous drone technology. At this rate, autonomous drones will inevitably surpass human piloting capabilities in the near future.

Direct Mapping Using a ‘Trained Expert’

The UZH autonomous racing drone flying through a forest without crashing
The UZH autonomous racing drone flying through a forest without crashing

Most autonomous drones rely on a series of processes to gain an understanding of their surroundings. These processes take time. The drone first collects data from its sensors, the onboard computer then processes that data, and finally maps it to create collision-free trajectories. Consumer drones such as Skydio rely on a similar computational model. However, the drones tested by UZH use a new method for obtaining autonomy at super-fast speeds.

Instead of processing the captured data, scientists at UZH directly mapped the noisy input from an Intel RealSense 435 stereo depth camera to create collision-free trajectories. This drastically reduced processing latency and provided results in real-time. Direct mapping works in an ingenious way. Scientists use a highly-trained neural network to accurately map the data coming from sensors.

The neural network is exclusively trained within a simulation using privileged learning. In privileged learning, the convolutional network has access to high-precision data (privileged data) such as 3D point clouds, perfect state estimation, and computation. These parameters are not available under real-life constraints. To train the network under real-life conditions, realistic sensor noise is simulated. This trains the network to make collision-free trajectories on realistic sensor data. Thus, when the network is tested under challenging real-world conditions onboard a drone, it outperforms existing traditional obstacle avoidance systems.

A process similar to privileged learning is used to train drone swarms in drone light shows. In a simulation, drones use 3D models to make collision-free trajectories and create visually stunning formations. Read more about how drone shows work here.

A Pilotless Future

The UZH team tested the drone in several challenging environments such as forests, buildings, disaster zones, and thick vegetation. The drone’s obstacle avoidance performed extremely well at high speeds. Such a level of obstacle avoidance at a speed as high as 40kph has never been achieved before.

While this system is not flawless, it sure is an upgrade over older systems. For instance, the direct mapping system does not function well in low-light scenarios. This can be fixed with a different camera setup. Additionally, as sensors improve, the latency and response time of autonomous drones will significantly go down.

“While humans require years to train, the AI, leveraging high-performance simulators, can reach comparable navigation abilities much faster, basically overnight.” -Antonio Loquercio, UZH, stated. Autonomy at high speeds would prove to be incredibly useful in disaster management, search and rescue, fire hazards, and even drone delivery in dense urban areas.

To learn more about UZH’s autonomous racing drone, check out their research paper here.