The robot's 244-milligram electronics package draws 120 milliwatts, about 6% of its power. Roughly 94% still goes to the wings through a wire.
MASSACHUSETTS: Researchers at MIT and Cornell University have flown a 1.29-gram flapping-wing robot that hovers and tracks paths using only its own sensors and computer, without the motion-capture cameras that have long guided insect-scale flyers. The work was published in Science Robotics on 18 November 2026. However, the robot still draws power through a wire, and a human operator steered its headline flight onto a sunflower.
Escaping the camera room
For more than a decade, insect-scale robots have depended on ceiling-mounted infrared cameras to know where they are. This vehicle instead carries a downward-facing optical-flow sensor, a time-of-flight distance sensor and an inertial measurement unit. A single microcontroller core processes their data in flight.
The team is led by first author Yi-Hsuan Hsiao, with senior investigators YuFeng (Kevin) Chen and Jonathan How of MIT and Elizabeth Farrell Helbling of Cornell. Notably, the robot’s predecessor from December 2025 could perform 10 flips in 11 seconds, but it relied entirely on external tracking and offboard computing. The new flyer trades that agility for self-contained perception.
Earning the payload
Carrying 244 milligrams of electronics posed the first hurdle. Enlarging soft artificial muscles usually reduces their power density, so extra size can fail to add lift. Accordingly, the team enlarged the actuators by 50% in length and radius while tuning wing and transmission geometry. That raised net lift by 46% and gave the loaded robot a lift-to-weight ratio of 1.9.
Vibration posed the second hurdle. The four wings beat roughly 330 times a second, according to secondary reporting, which can swamp delicate motion sensors. To counter this, the electronics sit on a vibration-damping fiberglass base.
Why speed beat precision
Flapping-wing robots are inherently unstable in pitch and roll, so heavy estimation algorithms add dangerous delay. How noted that a computer this small must handle both perception and control fast enough to keep the vehicle stable. Consequently, the team built a simplified, cascaded estimator that runs with the controller at 480 hertz. Its sheer update rate suppressed drift better than slower, more precise methods.
The flight logs from the February 2026 preprint show the trade-offs. Altitude, measured directly by the distance sensor, stayed within 0.25 centimetres root-mean-square error during a 5-centimetre hover. Lateral position, by contrast, drifted about 2 centimetres over 12 seconds. Moreover, lateral error rose 12.7% when hover height doubled to 10 centimetres, as the floor texture became harder to resolve.
Secondary reporting on the published paper puts hover position error at 65% higher than with motion capture. Even so, accuracy remained at the centimetre scale across hovers, figure-eights and short point-to-point flights.
The 6% power paradox
Computing is no longer the bottleneck. The electronics drew 120 milliwatts, about 6% of the power, while roughly 94% drove the wings. Furthermore, the researchers have identified ways to trim that draw, leaving headroom for a future forward-looking camera.
Propulsion is the harder problem. The soft muscles need about 2 watts and high driving voltages, both currently supplied through the tether. Rough calculations using typical lithium-polymer energy densities and packaging overheads suggest that a battery light enough to fit the robot’s spare lift might sustain only one to two minutes of real flight. Such a cell could weigh no more than about half a gram.
The sunflower caveat
The most striking demonstration needs careful reading. Outside the camera room, the robot flew for 30 seconds, rose over an obstacle and landed on a sunflower. Yet a human supplied high-level joystick commands, and technical reporting on the preprint notes that flight calculations for that trial ran on an offboard computer. Fully onboard computing was proven in the separate hover and trajectory tests.
The robot also has no forward camera. Therefore it cannot see flowers or obstacles ahead, and it would struggle in darkness, where optical flow loses its reference.
Bees still win outdoors
MIT points to precision pollination of crops such as almonds and blueberries, crop monitoring and inspection of tight industrial spaces. Chen himself has noted that natural bees excel in open fields, and he sees indoor farms as the more plausible niche. Similarly, breezes faster than the robot’s peak test speed of about 0.68 metres a second, or light rain, could overwhelm it outdoors.
What to watch next
Three milestones will mark progress. First, the team must fit a lightweight camera without slowing the 480-hertz loop. Second, it must develop a sub-gram battery and miniature high-voltage electronics to cut the wire. Third, it must show autonomous flower recognition and pollen transfer rather than human-guided landings.
Beyond the wire
The robot is neither untethered nor a replacement for bees. Ultimately, though, by closing its own stabilisation loop onboard, it moves the field’s main constraint from laboratory cameras to micro-power electronics and low-power vision, the next problems the team says it will tackle.
