Under ideal conditions, flying a quadcopter drone is easy. In fact, the design of these aerial vehicles makes them so stable that they practically fly themselves. But in the real world, ideal conditions are hard to come by. More often than not, gusts of wind and turbulent air make it very difficult to keep a drone under control, and that is bad news for everything from autonomous package delivery services to search and rescue operations that need an eye in the sky.
At present, drone control systems simply cannot handle everything that nature might throw their way. Things might generally go pretty well, but some situation will inevitably come along that was not accounted for by the developers of the algorithm, and that can spell disaster for the vehicle. That may no longer be the case in the future, however, if a trio of engineers at MIT has their way. They have been hard at work on a novel approach that enables drones to maintain stable flight under very difficult conditions — even conditions that had not been specifically planned for in advance.
An overview of the offline meta-learning and online adaptive control components (📷: S. Tang et al.)
Their method relies on a learning technique called meta-learning, which essentially teaches the system how to learn, and adapt, on the fly. It does this by replacing prior assumptions about the environment with learned models, and also by automating the selection of the best algorithm to respond to unexpected challenges. Traditional control systems often require engineers to guess in advance what kinds of environmental factors the drone may face. This guesswork is encoded into mathematical models, but those models can fall short when reality deviates from expectations.
Instead, the researchers built a neural network that can learn the behavior of these disturbances from just 15 minutes of flight data. And the system does not just learn from the data — it also decides how best to learn. It does this by selecting the most suitable optimization algorithm from a family of algorithms known as mirror descent. This is a significant upgrade over more conventional techniques that rely solely on gradient descent, which is just one member of the mirror descent family.
Simulations show the new controller (blue) has improved tracking accuracy (📷: S. Tang et al.)
A series of simulations and early experiments have shown that the new control method achieves a 50% reduction in trajectory tracking errors compared to existing baseline methods. And not only does the system keep drones on track more effectively, but its performance actually improves as conditions worsen. In stronger winds — the very situations where other control methods tend to fail — the new system continues to adapt and perform well.
The team is now working to test their system on real drones in outdoor environments. They are also exploring how the method could manage more complex scenarios, such as accounting for shifting payload weights or handling multiple simultaneous disturbances. With some refinement based on the outcome of these trials, this control system could keep fleets of drones safe and on course in the future.