This tiny drone with a tiny brain is smart enough to fly itself

An incredible new microchip let researchers build the smallest autonomous drone with an AI neural network running on just 1/100th of a watt.

This tiny drone with a tiny brain is smart enough to fly itself
The modified Crazyflie 2.0 drone is about four inches across. [Photo: courtesy of ETH Zurich]

European engineers say they’ve built the world’s smallest autonomous rotor drone—one that uses minimal battery power to run its AI. It’s a crucial development in the push to build fleets of self-navigating nano-drones (four inches across or less) that one day will carry everything from environmental sensors to tiny cameras for security or inspection duty.


The battery power issue is a big one for drones kept aloft by furiously spinning propellers. Flight times range from about five minutes for cheap toy models to 30 for big professional drones used in surveying or shooting movies. What researchers from the University of Bologna and the Swiss Federal Technology Institute (known as ETH Zurich) say they have done is essentially shrink the weight and power needs of their autonomous drone.

A research paper published earlier this month reveals they equipped a quadrotor nano drone popular with hackers, the $180 Crazyflie 2.0, with a camera, processor, and customized neural network that consumes less than 1/100th of 1 watt (94 milliwatts)—1% of the 27-gram (near 1-ounce) flyer’s meager power supply. That and the addition of a 5-gram circuit board for all the electronics had minimal effect on the drone’s battery life, they say.

“We did not test this extensively, but from first tests we saw that we were running out of battery after approximately five minutes,” wrote researcher Antonio Loquercio in an email to Fast Company. Crazyflie maker Bitcraze claims a seven-minute battery life [without the modifications], but Loquercio thinks that’s overly optimistic.

Battery life still has a long way to go, says Loquercio. “[It] would be nice to achieve 30 minutes of autonomy. This would be more than enough to inspect, for example, medium-size warehouses and get back to the charging station.” Despite the big gap to cover, at least adding autonomy has a negligible effect on battery life.

A tiny electric brain

The story now gets technical and acronym-heavy. But hang in there. It offers a fascinating, perhaps unsettling, preview of technology that will enable a near future of pervasive AI and digital monitoring that goes well beyond drones.

The Bologna-Zurich team’s autonomous nano drone uses a new mobile processor named GAP8. It packs eight processing cores (hence the name) optimized for running artificial intelligence applications, such as image recognition and analysis.


The custom circuit board built for the drone, called PULP Shield, weighs just 5 grams and consumes less than one watt. [Image: courtesy of ETH Zurich]
GAP8, in turn, is based on a new microchip architecture known as RISC-V. The term “RISC” may not sound familiar, but the technology is everywhere. First developed in the 80s, RISC has become the foundation for mobile processors and a key part of laptop and desktop CPUs. Short for “reduced instruction set computer,” RISC saves power by running lots of simple operations rather than a smaller set of complex ones. In March, its inventors received the Turing Award, nicknamed the “Nobel Prize of computing.”

RISC-V is the new, open-source version of the technology. The Bologna and ETH Zurich universities have been using it to develop ultra-low-power computing platforms, not just for drones, but for all types of connected gadgets that will make up the so-called “internet of things,” or IoT. “The efficiency and openness of RISC-V make it ideal for IoT,” said RISC co-inventor Dave Patterson in an email to Fast Company. “We’ll soon see many more such clever applications using RISC-V.”

Paring back the code

The researchers’ breakthrough was to squeeze an already lightweight neural network meant for big drones, called DroNet, into an even lighter form for the power-sipping chip. They first cut the precision of the network, using numbers accurate to fewer decimal places. Then they reduced the number of layers of virtual neurons that analyze images and the precision of the data that one layer passes to the next.

Finally, they tweaked the code to spread operations across the chip’s eight cores—using parallel processing to make up for the scarce memory and power. (High-end smartphone chips run about 15 times faster than GAP8.) Power savings also cut the network’s ability to process camera images from 20 to 12 frames per second. Despite all the compromises, the network still runs fast and accurate enough to recognize an obstacle and alert the drone in less than half a second. “This is more than [fast] enough for the speed at which the Crazyflie 2.0 can fly [four meters per second],” wrote Loquercio.

The AI modifications, contained on a circuit board called PULP-Shield, consume scarcely any of the drone’s power. [Image: courtesy of ETH Zurich]
Before the nano-drone could fly, the team had to custom-train their version of DroNet in real-world navigation. They captured video from cameras mounted on cars and bicycles and carried by a hiker. A powerful desktop computer crunched the data for about six hours to develop the slim code that was loaded into the Crazyflie’s brain. “Everybody with a decent desktop computer can repeat our results by using the source code we released,” writes Loquercio.

DroNet is limited to lateral movement, so the drone can move only from side to side to go around obstacles, not up and down to fly above or below them. “Sometime this year we plan to release a new version that overcomes this limitation,” wrote Loquercio.


The Bologna-Zurich team’s drone design won’t translate to real-world swarms immediately. “We still need to overcome many technical and algorithmic challenges before this will become a real product,” wrote Loquercio. “I believe that in the next 5 years we will start getting something working in controlled environments, [such] as industries or warehouses. To apply it ‘into the wild’ one should probably wait a bit more.”

But this project demonstrates that sophisticated AI can run on a tiny bit of silicon, on scarcely any power, virtually anywhere. The researchers say that their breakthroughs will apply not just to other drones, but to other robots, environmental sensors, security cameras, or other tiny gadgets that IoT boosters envision keeping a constant watch on our world. Encouraging or unsettling news, depending how you feel about AI systems monitoring our lives.

About the author

Sean Captain is a Bay Area technology, science, and policy journalist. Follow him on Twitter @seancaptain.