The Pulse
MIT Makes Tiny Flying Robot 447% Faster With AI
MIT researchers have developed an AI control system that lets an insect-scale flying robot move 447% faster and accelerate 255% more quickly than earlier versions. The tethered robot can complete 10 consecutive somersaults in 11 seconds whi
AI.info Team ·
MIT researchers have increased the speed of an insect-scale flying robot by 447 percent, giving the machine enough control to complete 10 consecutive somersaults in 11 seconds. The robot also accelerates 255 percent faster than the researchers’ previous best demonstration and stays within roughly 4 to 5 centimeters of its planned flight path.
The results, published in Science Advances, come from a two-part AI control system that combines a computationally demanding flight planner with a deep-learning model that can issue commands quickly enough for real-time control. The work addresses one of the central problems in flying microrobotics: small machines can be physically agile, but their controllers often cannot calculate the necessary corrections fast enough.
447 Percent More Speed From a New Control System
MIT’s robot is about the size of a microcassette and weighs less than a paperclip. Its larger flapping wings are powered by soft artificial muscles, but earlier versions relied on a controller tuned by hand. That limited how aggressively the robot could turn, pitch and recover from disturbances.
The new system begins with a model-predictive controller. It uses a mathematical model of the robot’s dynamics to plan difficult maneuvers while accounting for limits on force and torque. The researchers then use imitation learning to train a deep-learning policy on that planner’s behavior. The smaller model reproduces the planner’s decisions at much higher speed, allowing it to control the robot during flight.
Jonathan P. How, the Ford Professor of Engineering at MIT and a co-senior author of the paper, described the relationship between the hardware and the software in practical terms: “The hardware advances pushed the controller so there was more we could do on the software side, but at the same time, as the controller developed, there was more they could do with the hardware.”
Ten Flips Without Losing the Flight Path
Repeated flips present a demanding control problem. The robot must slow down at exactly the right moment so that each new somersault begins from the correct position and orientation. Small errors can accumulate across the sequence and send the machine into a crash.
In testing, the robot completed 10 somersaults in 11 seconds despite disturbances from wind. MIT also demonstrated a rapid movement called a saccade, in which the robot pitches sharply, moves to a new position and pitches in the opposite direction to stop. Insects use similar movements to reposition themselves and maintain a useful field of view.
“We want to be able to use these robots in scenarios that more traditional quadcopter robots would have trouble flying into, but that insects could navigate,” says Kevin Chen, an associate professor in MIT’s Department of Electrical Engineering and Computer Science, head of the Soft and Micro Robotics Laboratory and co-senior author of the paper. “Now, with our bioinspired control framework, the flight performance of our robot is comparable to insects in terms of speed, acceleration, and the pitching angle.”
The Controller Still Runs Outside the Robot
The demonstration does not yet show a fully autonomous flying insect robot. MIT says the controller currently runs on an external computer rather than onboard the machine, and the robot depends on a complex motion-capture system to provide position information.
That limitation matters because the proposed uses—searching earthquake rubble, moving through narrow spaces and operating around falling debris—would require the robot to sense its surroundings without external tracking equipment. The researchers plan to add onboard cameras and sensors so the machine can eventually operate outdoors.
Sarah Bergbreiter, a professor of mechanical engineering at Carnegie Mellon University who was not involved in the work, said the experiments are notable because the robot maintains control despite wind gusts, manufacturing variation and a power tether that can wrap around it during repeated flips. She also said the results suggest that less precise versions of the control policies could eventually run on the limited computing hardware available at insect scale.
From Laboratory Acrobatics to Search Missions
Small flying robots could enter gaps that are inaccessible to conventional quadcopters, but practical deployment will require more than speed. A useful search robot would need to carry sensors, avoid obstacles, maintain stable flight in changing air currents and communicate its position while operating with limited power.
MIT’s next steps include testing onboard sensing and studying whether multiple robots can avoid collisions and coordinate their movements. The current experiment establishes the flight-control result: a tethered, insect-scale machine can perform fast turns and repeated flips with an AI policy trained from a more computationally expensive planner. The robot still needs onboard perception before it can leave the motion-capture lab.
MIT News reported the work on December 3, 2025. The study is titled Aerobatic maneuvers in insect-scale flapping-wing aerial robots via deep-learned robust tube model predictive control.