Pushing the Speed Limit

It takes effort to learn to fly a drone well, but with a bit of practice just about anyone can become competent. But as the speed gets dialed up, drones become more difficult to handle. The aerodynamics become complicated, and instability is always lurking just around the next turn, which can mean expensive crashes. Not only is it difficult for a human to pilot a speeding drone, but it is also a challenge algorithmically for computers.

A team of engineers from MIT have taken on the challenge of developing an algorithm that will enable high-speed drones to find the fastest path to their goal, all while avoiding obstacles. Drones with this ability could be used in searching for survivors of natural disasters, or in quickly locating the source of a gas leak.

One of the challenges in developing an autonomous navigation algorithm for a drone is that it requires many experimental flights to be performed with real drones. Naturally, as these flights will need to examine the set of feasible trajectories, including those that are pushing the boundaries of what is considered safe, there are bound to be many crashes. These crashes are expensive, and the experimentation is time consuming, which limits the amount of data that can reasonably be collected.

The team from MIT largely circumvented these issues by developing their algorithm heavily with data obtained from simulated flights, with a smaller component involving real world flights. Moreover, the simulated flights helped the researchers to identify the flight plans that were most likely to be successful, so they were able to avoid running experiments that would be likely to damage their drones.

To test the algorithm, an obstacle course was created. Using flight plans chosen from the simulations, drones were taught to navigate the course. A conventional algorithm was also tested on the course for comparison. In all trials, the new algorithm completed the course with a quicker time — up to 20% quicker. The new method also made some surprising choices, at times intentionally slowing down around a corner, for example, in order to pick up speed later in the race.

The team is currently working to improve some aspects of their method to increase its applicability in the future. At present, the algorithm can only model obstacles in two dimensions, however they hope to change that by implementing some algorithms from the field of computer graphics. The researchers are also planning to perform more flights, and at higher speeds with more obstacles, which may allow the algorithm to be implemented in a wider array of circumstances eventually.