51动漫

THAT drone buzzing round your head might be smarter than you think. Small drones with neural hardware resembling brains will soon share airspace with other aircraft, seeing and avoiding potential hazards autonomously. The ability will help drones take on a host of new roles.

Big firms like Amazon, DHL and Google are developing their own drone fleets for rapid delivery of consumer goods, fast food and pharmaceuticals. However, current rules restrict drones to flying within visual range of a human operator because of the risk of collision. Drones need an automatic 鈥渟ense-and-avoid鈥� capacity before they will be able to make deliveries on their own.

Computers capable of recognising objects in video and responding in real time are too big and too power-hungry for small drones. That means drones have to rely on short-range sensors like radar, which may not give enough warning to avoid a collision.

The key may be to mimic how animal brains work; our brains are poor at number-crunching but can process complex sensory input faster than digital systems.

Bio Inspired Technologies of Boise, Idaho, is doing just that. It is building a sense-and-avoid system using a memristor, a resistor with a memory. Like the synapse in a biological brain, the memristor changes when impulses pass through it. Crucially, it is able to remember the impulse after it has stopped.

This capability forms the basis of a learning system that mimics neurons and the connections between them. A chip-sized neural system linked to the drone鈥檚 existing camera can be trained to recognise aircraft and other hazards at long range. Bio Inspired鈥檚 drone should be ready for its first flight later this year.

The system can also recognise objects like clouds, birds, buildings and radio towers, and uses visual cues to estimate how far away the objects are.

鈥淥bjects like other aircraft can be catalogued in a vague sense, meaning 鈥業 see an aircraft鈥�, or in an exact sense: 鈥業 see another drone鈥�,鈥� says Terry Gafron, CEO of Bio Inspired.

Equipped with this information, the drone plots a new flight path to avoid a hazard, updating it in real time as the threat moves.

鈥淣ature seems to use this approach very effectively,鈥� says David Warne of Queensland University of Technology in Australia, who has worked with artificial neural networks that let drones recognise vegetation.

Like others in this area, much of Bio Inspired鈥檚 research has been funded by the military. But it is likely that it will benefit the wider market. Sense-and-avoid will make it possible for fleets of small drones to criss-cross cities delivering packages. Like a bird or insect, a neural-enabled drone could fly to the trickiest landing place 鈥� even balconies.

聯Like a bird, a neural-enabled drone could fly to the trickiest landing place 鈥� even balconies聰

Being able to recognise objects autonomously will enable a range of applications for small drones. Some of these are in the area of precision agriculture.

鈥淭he crop drone is on everyone鈥檚 short list,鈥� says Gafron. Drones could survey a farm, recognise areas where crops aren鈥檛 thriving and move in for a closer view to establish whether the field needs water, fertiliser or fungicide.

In the industrial field, neural drones could patrol pipelines looking for leaks, or identify electrical faults on power lines.

Closer to home, smart drones could clean windows, pick up litter, clear gutters or weed your garden, or send information to your car about which parking spaces are open. 鈥淚t simply flies around town monitoring parking spaces,鈥� says Gafron.

Smart drones could even track animal populations, flying along livestock boundaries to track wolf populations for example. 鈥淣ot only could the system fly autonomously, but it could conceivably tell the difference between a deer and a wolf from the air,鈥� Gafron says.

Memristor-inspired drones are not the only approach. Last year, US agency DARPA unveiled the TrueNorth neural chip developed in conjunction with IBM. This is a simulation of a neural network using digital hardware with enough neurons to match agile flyers like bees.

This article appeared in print under the headline 鈥淎 drone that learns鈥�