for finding people under snow. The device emits a radio beacon a rescue team can pick up using another ARVA receiver device. The latter essentially operates as a direction finding device, generating a “U-turn” signal whenever it detects the person carrying it starts moving away from the emitter. Modern ARVA devices are able to reach a 5m accuracy in locating an emitter under 10m of snow. 20

Our goal is to control the drone so that it reaches the supposed location of an ARVA emitter. To that end, we integrate a Pieps DSP PRO20 ARVA receiver with the Pixhawk board. A custom PID controller aligns the drone’s yaw with the direction pointed by the on-board ARVA receiver. Roll and pitch, instead, are determined to fly at constant speed along the direction indicated by the ARVA receiver. Navigation is thus entirely determined by the ARVA inputs. We implement this controller both using reactive control by probing the ARVA device as fast as possible, and with time-triggered control at 400Hz, that is, the same as in the original time-triggered implementation of Ardupilot.

We place an ARVA transmitter at one end of Rugby, and set up the quadcopter at 100m distance facing opposite to it. Even though GPS does not provide any inputs for navigation, we use it to track the path until the first time the ARVA device generates the “U-turn” signal. We compare the duration and length of the flight when using reactive or time-triggered control. We repeat this experiment 20 times in comparable environmental conditions.

Results. Reactive control results in a 21% (11%) reduction in the duration (length) of the flight, on average. Time-triggered control also shows higher variance in the results, occasionally producing quite inefficient paths. Figure 9 shows an example. The path followed by reactive control appears fairly smooth. In contrast, time-triggered control shows a convoluted trajectory at about one-third of the distance, where the yaw is almost ± 90° compared to the target.

The logs we collect during the experiment indicate that the reason for this behavior is essentially the inability of time-triggered control to promptly react. Probably because of a sudden wind gust, at some point the drone gains a lateral momentum. Time-triggered control is unable to react fast enough; a higher than 400Hz rate would probably be

Figure 9. Example of ARVA-driven navigation when using reactive (black) and time-triggered control (yellow). Time-triggered control occasionally produces highly inefficient paths, whereas we never observe similar behaviors with reactive control.

needed in this case, and the drone turns almost 90°. We never observe this behavior with reactive control, which better manages available processing resources against environment influences.

6. CONCLUSION AND OUTLOOK

Reactive control replaces the traditional time-triggered implementation of drone autopilots by governing the execution of the control logic based on changes in the navigation sensors. This allows the system to dynamically adapt the control rate to varying environment dynamics. To that end, we conceived a probabilistic approach to trigger the execution of the control logic, a way to carefully regulate the control executions over time, and an efficient implementation on resource-constrained embedded hardware. The benefits provided by reactive control include higher accuracy in motion control and longer flight times.

We are currently working toward obtaining official certifications from the Italian civil aviation authority to fly drones running reactive control over public ground. Surprisingly, the major hampering factor is turning out not to be reactive control per se. The evidence we collected during our experiments, plus (i) additional fallback mechanisms we implement to switch back to time-triggered control in case of problems and (ii) extensive tests conducted by independent technicians and professional pilots, were sufficient to convince the authority on the efficient and dependable operation of reactive control.

Rather, the authority would like to obtain a precise specification of what kind of drone, intended in its physical parts, can support reactive control. In computing terms, this essentially means a specification of the target platform. This task represents a multi-disciplinary challenge, in that it requires skills and expertise beyond the computing domain and reaching into electronics, aeronautics, and mechanics. We believe much of the future of computer science rests here, at the confluence with other disciplines.

0

100

200

300

400

500

600

700

800

900

0 100 200 300 400 500 600

R ate at se c o nd sc a le ( Hz )

Time (s)
Reactive - Arch
Reactive - Lab

Traditional fixed-frequency

Figure 8. Average rate of control at second scale in two example Ardupilot runs. Reactive control adapts the rate of control executions both in the short and long term, and according to the perceived environment influence.