How a tiny flying robot learnt ten somersaults in eleven seconds

Ten somersaults in eleven seconds leave little time to recover between rotations. For MIT’s tiny flying robot, each rotation has to end in a state from which the next can begin. In the research published on 3 December 2025, solving that problem meant giving a powerful mathematical planner a much faster pupil.[2]

The 750-milligram robot spans four centimetres and uses four flapping-wing modules. In the reported ten-flip sequence, its root-mean-square position error was 2.49 centimetres; the paper describes the power wire wrapping around the body during the second and sixth flips. Motion-capture cameras measured the laboratory flights, while power and computation remained offboard.[1]

A body capable of more

Earlier in 2025, the team had redesigned the airframe. Its previous eight-wing arrangement allowed neighbouring wings to interfere with one another’s airflow. Four outward-pointing wings reduced that problem and freed space for future electronics.[3]

The artificial muscles were soft cylinders formed from elastomer layers and thin electrodes. Redesigned transmissions reduced buckling; longer wing hinges reduced twisting stress. That January design could hover for about 1,000 seconds. These were mechanical gains: stronger, more durable flight hardware gave the later controller more capability to use.[3]

The teacher does the expensive thinking

The December project combined two kinds of computation. A model-predictive controller used the robot’s dynamics to plan demanding trajectories while respecting available force and torque. Then imitation learning trained a neural policy to produce control commands. The expensive planner supplied examples; the smaller policy acted quickly enough for flight.[2]

The important training trick was to include states around the ideal trajectory. The researchers’ “tube” is a mathematical envelope of nearby motions under uncertainty. Training on that envelope gave the pupil examples of corrective actions, as well as examples of getting everything right.[1]

The main flight policy issued 1,000 control updates per second.[1]

Speed, wind and repeated flips were separate tests

On a figure-eight trajectory, maximum speed reached 1.97 metres per second. Separate repeated-saccade trials faced wind at 1.60 metres per second; across five such flights, root-mean-square position error was 4.58 centimetres. Those wind trials were distinct from the ten-flip demonstration.[1]

The broader engineering ambition is to reach spaces that larger aircraft struggle to enter. MIT’s researchers discuss future cameras and sensors, including flight patterns that might help a tiny robot see its surroundings. That remains a development goal.[2]

The controller gives the small airframe fast reflexes. The next engineering challenge is to give it the sensing, computing and power supply it would need to use those reflexes away from the laboratory.[1]

Report an error or suggest a correction