As with all flying machines, drones have several safety considerations that must be adequately addressed. One of them is landing the drone safely.

Landing Problem

Landing drones safely is an intricate endeavor due to certain factors. The complex turbulence generated by multiple rotors further compounds by the reflection of waves from the surface on which the drone is descending. The air turbulence, thus, generated is not fully understood. Hence, compensating for this phenomenon is not easy.

Take-off and landing are two of the most critical segments of drone flights for this very reason. Drones slowly descend while wobbling under the effects of turbulence. During final approach, the power shuts down and the drone falls before making ground contact. Indeed, a better system is needed for drone landing, one that can counter the effects of turbulence.

Neural Lander

Artificial intelligence experts and control engineers are collaborating at Caltech for the development of a system that leverages neural networks for landing drones quickly and safely. The ‘Neural Lander’ is the culmination of this endeavor. This system tracks the speed and position of the drone to learn about landing it safely. The system works by determining optimal rotor speeds and trajectories that will facilitate the smoothest descent possible. 

With this system in place, not only can drones land more safely, but they can also fly more smoothly midair even in the presence of sporadic gusts of wind. Another advantage of this system is that it will enable drones to consume less battery charge while landing. In this regard, a paper is being published at the International Conference on Robotics and Automation conducted under the auspices of the Institute of Electrical and Electronics Engineers (IEEE).

Deep Neural Networks

Deep neural networks are AI mechanisms that take inspiration from how the brain works. ‘Deep’ implies the extent to which data processes. The input data goes through several layers that process input data differently to generate increasingly complex output. Deep neural networks have the capability of learning autonomously, making them well suited for recurring processes. 

Researchers have deployed the spectral normalization technique to ensure smooth flight under the guidance of the deep neural network. The technique leads to seamless output that does not vary considerably, even in response to substantial fluctuations in the input. The result is tighter control and greater stability. 

Flight Improvements

The team achieved improved landings by studying deviation that occurred from ideal trajectories in a 3D environment. Three kinds of tests carried out to this end: straight vertical descent, landing in a curved line, and drone flight across broken surfaces. The last of these tests is the most complicated since the effects of turbulence are the most unpredictable with uneven surfaces. 

The results of this landing system are very encouraging. Vertical error reduces by 100%, while lateral drift goes down by 90%.