First preview of Amazon's new delivery drone – TechCrunch

For the first time, Amazon today introduced its latest fully electric delivery drone during its first conference re: March in Las Vegas. Chances are, it does not look like or fly like what you expect from a drone. However, it is an ingenious hexagonal hybrid design that has very few moving parts and uses the fairing that protects its blades as well as its wings as it moves from helicopter-like vertical flight to take-off. plane.

According to Amazon, these drones will begin to be delivered in the coming months, although we do not know exactly where this will happen.

Perhaps more importantly, the drone is packed with sensors and a suite of compute modules that perform a variety of machine learning models to ensure the safety of the drone. Today's announcement marks the first time that Amazon has publicly spoken about these internally designed visual, thermal and ultrasonic sensors, and how autonomous UAV flight systems maneuver it to at its landing point. The goal was to build a drone as safe as possible and able to operate independently. Even if he is not connected to a network and is facing a new situation, he will be able to react appropriately and safely.

When you see it flying in airplane mode, it looks a bit like a TIE fighter, where the kernel contains all the sensors and navigation technology, as well as the package. The new drone can travel 15 km and carry packages weighing up to five pounds.

This new design is quite different from previous models. I was lucky to see it before today's announcement and I admit that I was hoping for a much more conventional design – rather a refined version of the latest design, almost like a sled.

In addition to the cool factor of the drone, which is probably a little bigger than what you can expect, what Amazon really insists on this week is the suite of sensors and security features developed for the drone.

Before today's announcement, I met Gur Kimchi, Amazon's vice president for its Prime Air program, to talk about the company's progress in recent years and what makes this new drone special.

"Our meaning and avoiding technology is what makes the drone independently safe," he said. "I say independently, because it's different from other approaches where some of the safety features do not fall under the plane. In our case, they are in the plane. "

Kimchi also pointed out that Amazon had designed virtually the entire software and hardware stack of the drone internally. "We control the technologies of the aircraft, from raw materials to equipment, through software, structures, the factory, the supply chain and, possibly, delivery," he said. "And finally, the plane itself has unique controls and capabilities to react to the world."

What is clear is that the team tried to keep the actual flight surfaces as simple as possible. There are four standard aircraft control surfaces and six rotors. That's all. The autopilot, which evaluates all the sensor data and that Amazon has also developed internally, gives the drone six degrees of freedom of maneuver to get to the destination. The angular box in the center of the drone, which houses most of its capabilities and the packaging it delivers, does not rotate. He sits rigidly in the plane.

We do not know how bad the drone will be. Kimchi would only say that this largely meets the established safety standards and that the noise profile is also important. He compared that to the difference between hearing a dentist exercise and classical music. Be that as it may, the drone is probably strong enough to make it difficult to miss when approaching your garden.

To see what is going on around it, the new drone uses a number of sensors and auto learning models, operating independently of each other, that constantly monitor the camera. drone flight envelope (which, thanks to its unique shape and controls, is far more flexible than that of a regular drone) and environment. These include regular camera images and infrared cameras to get a view of its surroundings. There are many sensors on all sides of the plane so it can spot distant objects, such as an approaching aircraft, as well as nearby objects, for example when the drone lands.

The drone also uses various models of machine learning, for example to detect all air traffic around it and react accordingly, or to detect people in the landing zone or to see a line above it (This is a very difficult problem to solve, since the lines tend to be quite difficult to detect). To do this, the team uses photogrammetric models, segmentation models and neural networks. "We probably have advanced algorithms in all these areas," Kimchi said.

Whenever the drone detects an object or a person in the landing zone, it obviously fails – or at least delays – the delivery attempt.

"The most important thing the aircraft can do is make the right decision safely when exposed to an event that is not planned – for which it has never been programmed," said Kimchi.

The team also uses a technique known as Visual Simultaneous Localization and Mapping (VSLAM), which allows the drone to create a map of its current environment, even if it does not have prior information about a location or GPS information.

"This combination of perception and algorithmic diversity is what we believe makes our system particularly safe," said Kimchi. When the drone goes to the place of delivery or the warehouse, all the sensors and algorithms must always be in agreement. When one fails or detects a problem, the drone cancels the mission. "Every part of the system has to accept being able to continue," Kimchi said.

What Kimchi emphasized throughout our conversation is that Amazon's approach goes beyond redundancy, which is a pretty obvious concept in aviation and involves having multiple instances of the same embedded hardware . Kimchi argues that it is also important to have a diversity of sensors that are completely independent of each other. For example, the drone has only one impact sensor, but it also has several other ways to measure the same value.

However, Amazon is not yet ready to delve into the details of what the actual embedded hardware looks like. Kimchi said the system uses more than one operating system and a processor architecture, though.

It is the integration of all these sensors, the artificial intelligence and the very design of the drone that allows the whole thing to work. At some point, however, things will go wrong. The drone can easily handle a rotor that stops working, which is pretty standard nowadays. In some circumstances, it can even handle two failed units. And unlike most other drones, it can slide if necessary, like any other plane. But when he has to find a place to land, his intelligent intelligence goes into action and the drone will try to find a safe place to land away from people and objects – and he has to do it without having any prior knowledge of his environment.

To get to this point, the team actually used an artificial intelligence system to evaluate more than 50,000 different configurations. Only computational fluid dynamics simulations have required 30 million AWS computing hours (it's good to have a big cloud when you want to build a new, highly optimized drone, it seems). The team also performed millions of simulations, of course, with all the sensors, and examined all possible positions and ranges of sensors – and even different lenses for the cameras – in order to find an optimal solution. "The optimization is the right set of sensors and their configuration on the plane," Kimchi said. "You always have both redundancy and diversity, both in the physical domain – sonar versus photons – and the algorithmic domain."

The team has also executed thousands of hardware simulations in the loop in which all flight services are activated and all sensors perceive the simulated environment. Here too, Kimchi was not quite ready to reveal the secret sauce used by the team to do this work.

And the team has obviously tested drones in the real world to validate its models. "The analytical models, the computer models are very rich and very deep, but they are not calibrated in relation to the real world. The real world is the ultimate generator of random events, "he said.

It remains to be seen where the new drone will make its first deliveries. It's a secret that Amazon is not yet ready to reveal. This will happen in the coming months, however. Amazon started drone deliveries in England some time ago. So it's an obvious choice, but there is no reason for the company to be able to opt for another country as well. The United States seems to be an unlikely candidate, as regulations are still evolving, but perhaps this problem will be resolved by then as well. Anyway, what once looked like a Black Friday waterfall might land in your garden sooner than you think.