Four tiny wings. A very large asterisk.

A tiny flying robot makes a very good engineering detective story. The obvious suspects are weight, power, and the wings. But the arrangement of those wings matters too. In MIT’s January 2025 report, a redesign reduced the robot from eight wings to four, each pointing outward. Neighboring wings had been interfering with one another’s airflow.

Soft artificial muscles and revised transmissions supplied the motion while reducing mechanical strain. The lesson is satisfyingly physical: an assembly can behave differently from its individual parts. Four capable wing units still need to cooperate as one aircraft. Sometimes the useful design move is to give the working parts more room.

The January 2025 platform combined four single-wing modules. Dielectric elastomer actuators drove revised transmissions that reduced buckling and strain. The outward wing arrangement reduced aerodynamic interference. This hardware work addressed force generation and mechanical durability; later controller results should be identified separately.

The January study, published in Science Robotics, achieved a 1,000-second hover. Nearly seventeen minutes sounds like a battery specification until you ask what the experiment was measuring. Here, the useful result concerned sustained flight and durability; onboard batteries and sensors remained future work in MIT’s accompanying report.

Try reading a demonstration as a sentence with a few blank spaces: this machine did this task, for this long, with this supporting equipment. Filling those spaces makes the achievement easier to appreciate. Endurance of the mechanism and independence of the whole system are different questions. Both matter. They also invite different experiments, different fixes, and different reasons to be impressed.

The reported hover duration is a supported laboratory result. Treat it as evidence of sustained operation of the flight mechanism. Battery endurance requires an onboard energy source and its associated mass to be included in the tested configuration. The January article describes that integration as future work.

A separate August 2025 preprint tackled agile control on a 750-milligram platform. It reported lateral speed of 197 centimeters per second and ten consecutive flips in eleven seconds. Those moves depended on external motion-capture cameras, a control computer, and a power tether. Fast movement and outside support appeared together in the same experiment.

That is a useful way to look at a spectacular robot video: widen the frame in your mind. Where does its position information come from? Where are the commands calculated? What supplies the energy? These questions reveal the architecture behind the performance. A tiny airborne body can be the most visible component of a much larger machine.

The August 2025 controller study used external sensing, computation, and power. Wingbeat frequency stayed at 330 Hz while voltage commands changed wingbeat amplitudes. Tether forces were themselves disturbances. The reported speed and repeated flips demonstrate performance of this complete experimental setup, including its feedback system.

A February 2026 preprint moved another part of the system onto a different, 1.29-gram robot: sensing and computation. It also described an operator-directed flight outside motion capture that ended on a sunflower. Those demonstrations used different configurations; the landing does not establish that all computation was onboard for that flight. Onboard power remained a future goal. Keep the dates, bodies and configurations attached to the claims.

The sunflower makes a lovely destination, but the deeper story is the shifting boundary of the machine. A sensing package, a controller, a power supply, and a planner each solve a different problem. Watch those pieces move aboard one by one. That is how a laboratory flyer gradually becomes a more independent system.

The February 2026 study reports onboard sensing and computation on a 1.29-gram platform. Its operator-directed landing demonstration and fully onboard-compute configuration should not be assumed identical: most experiments used sensor boards without the MCU board. High-level planning and power autonomy remained open work. These results must stay separate from the 750-milligram aerobatics platform.