MIT's Radical Approach to Creating Safe and Reliable Autopilots for Flying

Alright people, imagine you're watching "Top Gun: Maverick" with Tom Cruise. Maverick's got a crazy task at hand—training young pilots to pull off a mission that seems downright impossible. These pilots have to fly their jets super low through a rocky canyon, all sneaky-like so they don't show up on radar, and then make a rapid climb without smashing into those canyon walls. Spoiler alert: With Maverick's help, those brave humans actually manage to accomplish the mission.

Now, here's the deal: Machines, on the other hand, would struggle big time with the same heart-pounding challenge. See, autonomous aircraft have a hard time figuring out the best way to reach their target without smacking into obstacles or getting detected. Most of the AI methods we have today can't solve this problem, known as the stabilize-avoid problem, and would fail to reach their goal safely.

But hold on to your seats, because the brains over at MIT have come up with a new technique that beats the rest. These smarty-pants researchers have developed a machine-learning approach that can handle those complex stabilize-avoid problems like a pro. Not only does it match or even surpass the safety of other methods, but it also brings a whopping tenfold increase in stability. That means the machine can reach its goal and stay there without wobbling all over the place.

In an experiment that would make Maverick proud, their technique managed to pilot a simulated jet aircraft through a narrow corridor without a single crash. Talk about impressive!

Chuchu Fan, the Wilson Assistant Professor of Aeronautics and Astronautics, and the senior author of the research paper, explains, "This problem has been a tough nut to crack. Lots of people tried to tackle it, but they couldn't handle the high complexity and dynamics involved."

The MIT researchers took a fresh approach to the problem. First, they broke it down into two steps. Step one: they transformed the stabilize-avoid problem into an optimization problem with constraints. This allowed the machine to reach its goal while making sure it dodged obstacles. Then, for step two, they used a deep reinforcement learning algorithm to solve this optimized problem in a mathematical form known as the epigraph form. This approach helped them overcome the hurdles faced by other methods.

But it wasn't a walk in the park. See, the deep reinforcement learning algorithm wasn't initially cut out for this kind of optimization problem. The researchers had to come up with new mathematical expressions that fit their system. To make things work, they also combined these new tricks with some engineering techniques used by other methods.

So, how did their technique hold up in the real test? Well, they set up various control experiments with different starting conditions. In some simulations, the autonomous agent had to reach and stay inside a goal region while pulling off intense maneuvers to avoid obstacles coming right at them.

When they compared their approach with several other methods, theirs was the only one that could stabilize all the trajectories while ensuring safety. And just to push their limits, they used their technique to fly a simulated jet aircraft in a scenario straight out of a "Top Gun" movie. The jet had to stabilize near the ground, maintain a low altitude, and stay within a narrow flight corridor.

Now, this simulated jet model was actually released to the public back in 2018, and it was designed as a challenge for flight control experts. The researchers wondered if their controller could handle a scenario that would bring it down. Turns out, the model was a real tough nut to crack and couldn't handle complex scenarios. But guess what? The MIT researchers' controller aced it! It prevented the jet from crashing or stalling and did a way better job at stabilizing towards the goal compared to the other methods.

Looking ahead, this technique could be a game-changer for designing controllers for super dynamic robots that need to meet safety and stability requirements. Think about autonomous delivery drones, for example. It could also be integrated into larger systems. Picture this: you're driving on a snowy road, and your car starts sliding. The algorithm could kick in and help you steer your way back to safety.

The real strength of this approach lies in navigating extreme scenarios where a human would struggle. Oswin So, the graduate student and lead author of the paper, says, "Our goal as a field should be to make reinforcement learning safe and stable when we use it in critical systems. We see this as a promising first step towards that goal."

Moving forward, the researchers want to make their technique even better at dealing with uncertainty in solving optimization problems. They also plan to test how well the algorithm works in the real world, on actual hardware, considering that there might be differences between the model dynamics and real-life situations.

Stanley Bak, an assistant professor at Stony Brook University, who wasn't involved in this research, is pretty impressed. He says, "Professor Fan's team has improved reinforcement learning performance for dynamical systems where safety matters. They're not just aiming for the goal; they're creating controllers that ensure the system can reach its target safely and stay there indefinitely."

This awesome research is funded in part by MIT Lincoln Laboratory, as part of their Safety in Aerobatic Flight Regimes program. So, with the MIT team leading the way, it seems like we're getting closer to a future where machines can handle crazy challenges with grace and make our lives easier and safer.