A one-and-a-half-meter-long helicopter sits on a table in the basement of Building 33. Instead of a sleek fiberglass shell, its body is a roughly soldered metal framework with the “finish” of an Erector set. It is utilitarian, not ornamental, but what it lacks in style, it makes up for in substance. This minicopter can do things larger helicopters can’t. And even more important, it can do them on autopilot. It is essentially a flying robot. Apart from takeoff and landing, all its movements are controlled by mathematical equations.
Last fall the machine, dubbed Mr. Chopper, broke new ground when it became the first helicopter to perform a split-S maneuver-a half roll followed by a half loop-without human intervention. The maneuver capped a host of accomplishments by a research team from the Departments of Aeronautics and Astronautics and Electrical Engineering and Computer Science. Led by aeronautics and astronautics associate professor Eric Feron, the team is focused on enhancing aircraft agility so that flying machines can do loops and turns in tight spaces. Members say that in the next 10 years, nimble, unmanned vehicles like Mr. Chopper could be used for military reconnaissance and filming movies.
“I think it’s probably the craziest flying project throughout the country in the university environment,” says Feron, who began working with minicopters in 1998 after watching the pilot of a remote-control helicopter direct a chopper through several stunts. Previously, Feron had studied automatic control systems, and the idea of automating helicopter flight intrigued him. He enlisted two graduate students, Alex Shterenberg, MNG ‘00, and aeronautics and astronautics PhD candidate Vlad Gavrilets, SM ‘98, to build the avionics box that now directs Mr. Chopper’s movements.
The four-and-a-half-kilogram avionics box acts as the controller’s eyes in the sky by continuously measuring and transmitting flight data to the ground. Attached below the helicopter’s three-kilogram body, the box contains three sensors, a Global Positioning System receiver, an altimeter, and a flight control computer. It communicates via Ethernet with a ground computer, which the graduate students also built.
“What Vlad and his coworkers have done is literally build an entire computer system, sitting underneath the helicopter, from basic components,” Feron says. “It has to work in an environment that shakes a lot and not only computes things, but also sends orders to physical devices.”
To gather the data that direct the helicopter during automated flight, Shterenberg and Gavrilets outfitted the craft with a custom-built data-acquisition box. Remote-control helicopter pilot Raja Bortcosh led the chopper through myriad maneuvers, and the box recorded his commands and the helicopter’s sensor outputs. “Using that data,” Gavrilets says, “we were able to reconstruct the way the pilot directs the helicopter to perform the maneuvers. And then we were able to build the first dynamic model-mathematical model-of the helicopter in aerobatic flight, which had never been done before.” Once they had reconstructed the flight commands as mathematical equations, the researchers-joined by postdoctoral associate Bernard Mettler, aeronautics and astronautics graduate student Ioannis Martinos, and electrical engineering and computer science graduate student Kara Sprague ‘01, MEng ‘02-programmed the information into Mr. Chopper’s computer, allowing the machine to duplicate the maneuvers on its own.
Mettler, whose doctoral research at Carnegie Mellon University focused on modeling and control techniques for miniature helicopters, notes that it takes years to become proficient at controlling a helicopter, so it’s not easy to duplicate these skills with a computer. “The MIT autonomous helicopter group, by successfully executing aerobatic maneuvers, achieved a new state of the art in flight performance under computer control,” he says.
To test their mathematical models, the researchers built a flight simulator with a duplicate avionics box and interface. Through the simulator, they can watch a 3-D image of a helicopter move wherever their models direct it. The team tries to eliminate all mathematical inaccuracies on the simulator before it tests the commands with the helicopter.
On the field, with the selected mathematical algorithm programmed into the helicopter, the pilot controls the chopper as it takes off from the ground. Directing it to hover at a specific location and altitude, the pilot flips a switch that puts the helicopter on autopilot. The helicopter follows the instructions of the algorithm and performs the maneuver within prescribed parameters of altitude, speed, and distance, and all the while, the researchers monitor its flight data on the ground computer. After completing the maneuver, the chopper returns to its hovering position, and the pilot takes manual control and lands the chopper.
Thanks to the simulator and the researchers’ ability to test the mathematical models before taking to the skies, the helicopter has crashed only twice in its four-year history, despite the fact that, as Feron says, “the windows for making mistakes and recovering from them are extremely narrow.” Both crashes were attributable to hardware failures, not numerical errors.
But last fall team members observed an unexpected “test” of emergency recovery procedures, the programmed instructions that direct the helicopter to respond to occasionally erroneous sensor data. They were trying out an algorithm, which up until then had been tested only on the simulator. The model was to direct the helicopter to perform two automatic maneuvers in a row, an aileron roll-a corkscrew-like maneuver-followed by a split-S. But when it came time to actually execute the maneuvers, there was a problem: as the chopper was coming out of the roll, it received faulty data about its position, and it responded by going into a spiral dive. Instead of attempting to take over and rescue the helicopter, the team waited to see whether the craft’s sensors would receive correct data upon which it could act. To their delight, that is exactly what happened. Unaided, the helicopter recovered from the dive and returned to its starting position.
One of the most dramatic flight moments came when the helicopter first performed a split-S. The researchers had predicted it would drop 36.5 meters in 2.5 seconds as it came out of the split-S, and it did. But the drop-similar to an error-induced free fall-was more nerve-wracking than anticipated. “On a simulator, it doesn’t look nearly as dramatic as when you actually see it,” Gavrilets says. “It was really wild!”
With its ability to perform tight turns, loops, and rolls, the miniature helicopter is perfectly suited for negotiating urban and natural landscapes, Mettler says. Already it has been used to help film a movie in New York City, although for that, Mr. Chopper operated under human control. A camera attached to its base recorded pigeons in flight for a 2000 Emmy-winning documentary by National Geographic. The helicopter’s built-in vibration-isolation system, which cushions the avionics computer during flight, allowed for even filming. The craft’s steady-shot capability and its maneuverability are attracting entertainment industry interest because the machine could be used to film stunts inexpensively.
But Feron and his team most often discuss military applications. After all, NASA, the Office of Naval Research, and the Defense Advanced Research Projects Agency fund the project. Last fall the team developed an “air show” program designed to demonstrate that the helicopter can autonomously carry out a planned mission-an ability necessary for such military uses as reconnaissance.
Related work (on larger helicopters) is going on at other universities, including Carnegie Mellon and the Georgia Institute of Technology, but researchers there recognize MIT’s advances. Georgia Tech assistant professor Eric Johnson, SM ‘95, says, “In my view, the impact of Eric Feron and his team’s recent results has been to unambiguously show that unmanned aerial vehicles are capable of operating more like manned aircraft.”
This new data poisoning tool lets artists fight back against generative AI
The tool, called Nightshade, messes up training data in ways that could cause serious damage to image-generating AI models.
The Biggest Questions: What is death?
New neuroscience is challenging our understanding of the dying process—bringing opportunities for the living.
Rogue superintelligence and merging with machines: Inside the mind of OpenAI’s chief scientist
An exclusive conversation with Ilya Sutskever on his fears for the future of AI and why they’ve made him change the focus of his life’s work.
How to fix the internet
If we want online discourse to improve, we need to move beyond the big platforms.
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.