A drone from Zipline travels at 70 mph and the ideal situation is for no other low-flying aircraft to crash into it. Zipline drones are used to deliver medical supplies, such as blood, via parachute in countries like Rwanda and Ghana. They also deliver commercially from an Arkansas Walmart. They have even delivered semen from bulls and pigs to Rwanda for insemination purposes, with a particular focus on genetic diversity and milk production in the case the bull semen.
Zipline is like any other operator of aircraft, and doesn’t want to be in midair collisions. Keenan Wyrobek is the company’s chief technology officer. He says it can be difficult to navigate the airspace in the US. Regular people flying low-flying Cessnas, or those operating crop dusters, may not have a transponder to announce their location. Wyrobek states that transponders are not mandatory for many of these planes. “There is a wild west spirit in aviation in this country.”
A human pilot of a small plane must look ahead to avoid hitting another aircraft, such as a hovering helicopter. However, the UAV must take the necessary actions to avoid being struck by itself. Wyrobek states that it is the drone’s job to spot a Cessna approaching them and move out of their way. In this case, the poky cowboy must make way for the cowhand galloping much faster.
This raises questions that are not limited to this company. How can a drone without a human operator to monitor traffic, avoid collisions, and stay out of harm’s way of a fast-flying aircraft? How does it detect threatening traffic?
Zipline’s solution to the larger problem does not include radar, cameras or lidar. These are common approaches in autonomous car spaces. The drone company decided to use microphones that can detect other aircraft and allow the drone to get out of harm’s way.
This sensor array is composed of eight microphones that are placed on probes protruding from the 11-foot wings. The system must be able ignore ambient noise from the drone (the drone’s own propeller sounds) and just listen out for other flying machines. Wyrobek states that the array is essential to help get enough signal-to noise to hear planes far away and also to find them. It can then “triangle where these planes are actually coming from.”
The drone uses a very small amount of computing power to perform this trick. He says that the drone uses a mix of signal processing techniques, such as beam forming, and machine learning, AI-based techniques to locate where it is. This job is made easier by an onboard GPU, which can handle AI-related tasks. Wyrobek states that the microphones do not produce much data. Wyrobek adds that the actual information pipe is small so it doesn’t have a large compute load.
They collected training data, which included approximately 15,000 planned interactions between drones and human-crewed aircraft such as an airplane or helicopter. Although microphones won’t be very helpful for a hot air balloon, or glider, Wyrobek states that this approach has a fortunate benefit: faster-moving aircraft tend to be louder which means that the signal from a faster-moving threat will be stronger.
The company is currently waiting on regulatory approval before the drone’s software can make decisions about whether to take evasive actions to avoid any aircraft that might strike it. This would include a maneuver like proactive turning away from the way and entering an awaiting pattern until the coast clears. The microphones are currently installed in certain drones even though the entire system has not yet been switched on. He says that the microphones believe they are in control. However, the team reviews the data post-flight to ensure the system does what it was supposed to.
Wyrobek expects to use the sound-based detection software and avoidance software in all regions where they work. He says that he thinks it will be used in most areas. “As we scale up, we want to continue increasing our safety and this is one way to do so.” This sounds great from an airspace management perspective.
