Scientists from Delft University of Technology's (TU DELFT) use Drone races to test AI based on neural network intended for future space missions. This innovative research was cooperation between the advanced team of the concept of the European Space Agency and Micro Air Vehicle Laboratory (Mavlab) in TU Delft.
The project aims to examine the use of neural networks to autonomous management of complex spacecraft maneuvers, such as interplanetary transfers, surface landings and docking. In a difficult environment of space, maximizing the efficiency of on -board resources, including propeleant, energy, computing force and time, is crucial. Neuron networks can optimize on -board operations, increasing both mission autonomy and reliability.
To confirm these neural networks in real conditions, scientists chose drone races as perfect tests. Cyber Zoo, a 10 x 10 meters testing area at the Faculty of Aviation Engineering TU Delft, provided a perfect setting. Here, drones with people alternately with autonomous drones equipped with neural network trained in various methods.
Racing drone serves as excellent test tests for comprehensive neural architecture on real robotic platforms, helping researchers to build trust in their ability to apply them to space missions. Drones raced through the established course, simulating the restrictions and challenges that a spaceship would encounter during the mission.
Traditionally, a spacecraft maneuvers are carefully planned on Earth, and then sent to a spaceship. Some tips take place on the ground, while the control part is served by a spaceship. However, the unpredictable nature of space, with variables, such as gravitational shifts and atmospheric turbulence, is serious challenges.
An alternative approach, called comprehensive tip and control networks (G&C networks), includes all processes taking place on a spaceship. Instead of observing a predetermined course, the spacecraft constantly repels the optimal trajectory from the current position, which causes much greater performance. This method drastically reduces the costs of resource related to traditional corrections of brutal force.
There are many similarities between drones and a spacecraft, although the dynamics of drones is faster and louder. In races, time is the main limitation, but it can be used as proxy for other key variables for space missions, such as fuel mass.
Despite the limitations of satellite processors of G&D networks, they are surprisingly compact, storing up to 30,000 parameters in memory with only a few hundred kilobytes and covering less than 360 neurons.
In order for G&C networks to be effective, they must send commands directly to the cylinders – a gun for a spaceship and a propeller for drones. The main challenge was the solution to the gap in reality between the simulated and real cylinders. The team coped with this by teaching neural networks to adapt to real conditions. For example, if the propellers provide less pushing than expected, the drone detects with its acceleraometry, which prompted the neural network to adapt commands to follow the new optimal path.
The living academic community in drone races is a great opportunity to test and improve these systems. Using drones helps to determine the solid theoretical framework and set safety parameters before planning the actual space mission.
In the zoo cyberspace, they not only compete for speed, but also pave the way for future space exploration. By placing AI control systems based on a neural network in this demanding environment, ESA and TU Delft make significant progress towards more autonomous and efficient cosmic missions.
