MIT Breakthrough Brings Delivery Drones Closer To Reality
Eric Hopton for redOrbit.com – Your Universe Online
The much vaunted drone delivery promise has taken a step closer to reality.
In December last year the online retail giant Amazon revealed its “Prime Air” plans to deliver its products by drone – the so called “Octocopter” which would, it was claimed, drop your latest purchase at your door in as little as thirty minutes.
Amazon had kept their brainchild under close wraps until the filming of a “60 Minutes” interview with Amazon top man Jeff Bezos. Though Amazon had dropped giant hints that they were about to reveal a top secret project, they gave away no clues. During a break in filming, Bezos taunted the producers and told them “If you can guess what it is, then…I will give away half my fortune and send you to Vegas with it.” Bezos kept his millions and the crew never got their trip to Vegas.
CBS News reported what happened after the documentary was in the can. The film crew was taken into a “mystery room” and saw what the fuss was all about. They were shown what looked like “giant, flying tarantulas” – the Octocopter Drones – and a demo video of how the drones would do their job.
Many experts at that time questioned whether the software needed to accomplish this was available or even possible. The sheer number of variables involved – from weather to navigation, sensor, fuel, and mechanical issues – made this seem like a very big task even though the likes of Amazon and Google were throwing their best minds at it.
Researchers at MIT have taken on the technical challenge and claim to have developed a “two-pronged approach that significantly reduces the computation associated with lengthy delivery missions.”
The team at MIT first tackled the problem of computing what they term the “health” of the drone. They developed an algorithm that can measure and predict, in real time, the drone’s fuel levels as well the condition of essential equipment such as the propellers and the cameras and other sensors which enable it to fly and navigate accurately. This is a sophisticated process and even has the capability to instruct the drone to take any necessary proactive measures like changing its programmed route to call in at a charging station.
Simulating scenarios where the drones make multiple deliveries in “various environmental conditions,” the researchers discovered that the machines fitted with the advanced “health” monitoring technology were every bit as efficient as the older models but were much more reliable. They were far less prone to failures and breakdowns. Ali-akbar Agha-mohammadi of the MIT Department of Aeronautics and Astronautics said that system health is essential when drones are making multiple drops over long hours in harsh environments. The new algorithm, he claims, resulted in only a few failures over the course of 100 tests.
The second part of the “two-pronged” strategy was to improve the navigational capabilities of the drones. They did this by building in a system that allows the drone to “efficiently compute its possible future locations offline, before it takes off.” The drone can then simplify all the possible routes to its destination and avoid any obstacles.
Planning routes for autonomous vehicles is highly complex. A common approach is to use a computational system known as the Markov Decision Process (MDP). This is a “sequential decision-making framework that resembles a “tree” of possible actions.” At each branching point of the tree new potential actions are opened up. Every subsequent “node” presents other possibilities so the number of choices grows rapidly. The role of MDP is “the process of reasoning about the future,” making the appropriate choices and reducing risk.
In simulated environments with perfect measurements MDP gives excellent results. In the real world where the drones will operate there are few perfect measurements. The MIT team used the example of an instruction to turn at an angle of 90 degrees. This presents no problems in a simulation, but out in the field a strong wind will cause the drone to deviate and make computation that much harder.
So the research team chose to employ a “more general framework” – the Partially Observable Markov Decision Processes (POMDP). This reduces the huge number of choices in the “tree of possibilities” and therefore “funnels multiple possible outcomes into a few most-likely outcomes.” Plans are now in place to run tests in actual real environment situations.
Agha-mohammadi co-authored a report on the advanced drone system with MIT graduate student Kemal Ure, Professor of Aeronautics and Astronautics Jonathan How, and Boeing’s John Vian. The findings will be presented at the Chicago IEEE/RSJ International Conference on Intelligent Robots and Systems in September.
> Further Reading: The History Of Drone Technology