A new algorithm will allow drones to be swift and agile for urgent operations like search and rescue.
You will likely recall the crashes and the victories if you are a follower of autonomous drone racing. Teams compete in drone racing to determine which drone can fly the fastest through obstacles. The faster drones fly, however, the more unstable they are. At high speeds, their aerodynamics can prove difficult to predict. Crashes are often very spectacular and ordinary.
Drones can also be used in critical operations outside of the race course. For instance, to search for survivors after a natural catastrophe.
MIT aerospace engineers have created an algorithm to help drones navigate around obstacles and avoid crashing. This new algorithm combines simulations of drones flying through virtual obstacles and data from real drones flying through the same course.
Researchers found that a drone equipped with the algorithm could fly through an obstacle course 20 percent faster than a drone using conventional planning algorithms. The new algorithm only sometimes kept the drone ahead of its rivals throughout the course. It could slow down a drone to deal with a complex curve or save energy to speed up to overtake its competitor.
Ezra Tal is a graduate student at MIT’s Department of Aeronautics and Astronautics. “At high speed, intricate aerodynamics are difficult to simulate. So we use experiments in real life to fill in those black holes to find, for example, that it might be more beneficial to slow down first than to be faster later.” We use this holistic approach to determine how we can make a trajectory overall so fast.
“These algorithms are a beneficial step towards enabling future drones to navigate complex environments very quickly,” Sertac Karaman (associate professor of aeronautics and Astronautics and director at the Laboratory for Information and Decision Systems, MIT) said. “We hope to push the limits so they can travel as fast and far as their physical limits permit.”
Fast effects
If drones are designed to fly slowly, it is easy to train them to navigate around obstacles. Because aerodynamics like drag is only sometimes relevant at low speeds, they can be left out in any modeling of drone behavior. These effects are more apparent at higher rates, and predicting how the vehicles will react is harder.
Ryou states, “When you fly fast, it can be difficult to know where you are.” There could be delays in signaling to the motor or a sudden voltage drop that could lead to other dynamics problems. These effects cannot be modeled using traditional planning methods.
Researchers have conducted many experiments in the laboratory to understand how high-speed aerodynamics affects drones in flight. They set drones at different speeds and trajectories to determine which drones fly fast without falling. This is a costly and sometimes dangerous training method.
Instead, the MIT team created a high-speed flight planning algorithm that combines simulations with experiments to reduce the number required for safe and fast flight paths.
Researchers started with a physics-based flight plan model. This was used to simulate the behavior of a drone while flying through an obstacle course. They created thousands of simulations, each with a different speed pattern and flight path. Then they charted which scenarios were feasible (safe) and which resulted in crashes. They were able to quickly identify the best possible methods or racing trajectories to test out in the lab.
“We can quickly and cheaply simulate low-fidelity and see interesting trajectories. This could make it feasible to fly them fast. Tal explains that we then fly these trajectories in experiments to verify their feasibility in the real world. “Eventually, we find the best trajectory for us in the shortest time possible.”
Slow to move fast
The researchers created a simulation of a drone that flew through a course that had five enormous square-shaped obstacles. They arranged them in a staggered arrangement to demonstrate their new approach. This same course was set up in a training area. The researchers then programmed the drone to fly through it at speeds and trajectories they had previously identified from simulations. The same course was also run with a drone trained using a conventional algorithm and did not include experiments in its planning.
The drone trained on the new algorithm won every race and completed the course in less time than the traditionally trained drone. In some cases, the winning drone ended the study 20 percent faster than the competitor, although it had a slower start and took more time to bank around turns. The conventionally trained drone did not make this subtle adjustment, probably because the simulations used to calculate the trajectory could not fully account for the aerodynamic effects observed in the real world.
To improve their algorithm, the researchers will fly more experiments at higher speeds and in more complex environments. They may also use flight data from remote pilots, whose decisions and maneuvers could help them pinpoint even more feasible flight plans.
Tal states that if a human pilot slows down or picks up speed, it could help inform our algorithm’s actions. We can also use the human pilot’s trajectory as a starting point and work from there to determine if our algorithm can fly faster. These are future ideas that we are considering.
