MIT engineers develop algorithm to train drones to avoid obstacles at high speeds
Anyone who follows autonomous drone racing has probably seen more crashes than wins. In drone races, teams compete to determine which device is best prepared to fly through an obstacle course. However, the faster drones fly, the less stable they become. At high speeds, their aerodynamics are difficult to predict. Crashes are a common occurrence.
If drones could be made faster and more maneuverable, they could be used in time-critical operations beyond the racetrack, for example, to search for survivors after a natural disaster.
Aerospace engineers at the Massachusetts Institute of Technology (MIT) have developed an algorithm that helps drones find the fastest route around obstacles without crashing. It combines simulations of a drone flying through a virtual obstacle course with data from experiments of a real drone flying the same route in physical space.
The researchers found that a drone trained with their algorithm flew through the obstacle course 20% faster than a drone trained with conventional planning algorithms. However, the new algorithm did not always allow drones to outpace their competitors throughout the race. Sometimes it chose to slow the drone to take a tricky turn, or to save energy so it could accelerate and ultimately overtake its rival.
“At high speeds, complex aerodynamics arise that are hard to model, so we use real-world experiments to fill these black holes and find out, for example, what may be better to slow down first and then accelerate. It is this holistic approach that we use to see how we can make the trajectory as fast as possible overall”
According to Sertac Karaman, MIT associate professor of aeronautics, such algorithms are a very valuable step toward building the drones of the future that will be able to move very quickly in complex conditions: "We really hope to push the boundaries so they can travel as fast as their physical capabilities allow.
Fast effects
Training drones to avoid obstacles is relatively simple if they must fly slowly. This is because aerodynamics, such as drag, are usually not significant at low speeds and can be ignored when modeling drone behavior. But at high speeds such effects are much more pronounced, and it is far harder to predict how the devices will cope.
During fast flight, it is difficult for a drone to estimate where it is. There may be delays in sending a signal to the motor or a sudden voltage drop that can cause other dynamics problems. These effects cannot be modeled using traditional planning approaches.
To understand how high-speed aerodynamics affect drones in flight, researchers have to run many experiments in the lab, tuning drones at different speeds and trajectories to see which fly fast and without crashing. This is an expensive and often crash-prone process.
Instead, the MIT team developed a high-speed flight planning algorithm that combines simulation and experiments in a way that minimizes the number of experiments needed to identify fast and safe flight trajectories.
The researchers started with a physics-based flight planning model, which they developed to first simulate the drone's behavior while flying through a virtual obstacle course. They simulated thousands of racing scenarios, each with its own flight trajectory and speed.
They then marked whether each scenario was feasible (safe) or infeasible (leading to a crash). From this table, the researchers were able to quickly select several promising scenarios or racing trajectories to try in the lab.
“We can run this simulation quickly and cheaply to see interesting trajectories that may be fast and realistic. Then we run experiments on these trajectories to understand which of them are actually possible in the real world. Ultimately, we approach an optimal trajectory that gives us the lowest possible time”
Slowing down to speed up later
To demonstrate their new approach, the researchers simulated an unmanned aerial vehicle flying along a simple route with five large square obstacles arranged in a checkerboard pattern.
They set up the same configuration in an indoor physical training space and programmed the drone to fly the course at the speeds and trajectories they had previously selected from their simulations. They also ran the same course with a drone trained with a more traditional algorithm that does not include experiments in its planning.
Overall, the drone trained with the new algorithm "won" every race, completing the course in less time than the conventionally trained drone. In some scenarios, the winning drone finished the race 20% faster than its competitor, even when it chose a trajectory with a slower start, for example, taking a little more time to turn.
The researchers plan to run more experiments at higher speeds and in more complex conditions to improve their algorithm. They may also incorporate flight data from human pilots who remotely control drones, whose decisions and maneuvers could help achieve even faster yet still feasible flight plans.
