Interesting video from Microsoft’s AI research team on using AI to find thermals. You can see Mission planner being used at the 0:46 mark.
From the video it looks like it could be just arduplane code that has thermaling build in as well these days.
Would be interesting to know how smart their ai really is, how well it worked etc.
I would like to say a huge – HUGE – thanks to this project and MissionPlanner.
@Linus is correct – we have been using MissionPlanner, and our autopilot software is based on ArduPlane. In particular, we are using a heavily modified version of the soaring controller (of the AP_Soaring library, to be specific). Its smarts come from lean variants of AI techniques such as Bayesian reinforcement learning (BRL). More details will follow in conference publications. As we’ve come to realize though, even stripped-down versions of BRL are too resource-heavy for something like a Pixhawk, and ultimately don’t need to be run on dedicated AutoPilot hardware. The plan is to port the AI part of the controller to Raspberry Pi 3, but still rely on modified ArduPlane for other functionality.
Microsoft is using the autonomous soaring capability in ArduPlane
Might be a good use case for the Intel Movidius AI Stick ?
lol… I think this concept is ‘‘in the air’’ :
I like the rpi3 as companion computer
Ostensibly, there are many options for a companion computer, but many of them are optimized for neural networks. In our case, this is unnecessary at this point. We just need a capable general-purpose board with – and this is really important – low power consumption. Few alternatives can beat RPi3 from that standpoint.
What is the airframe you are using?
We are using Radian Pros, a Thermik XXXL, and an F3J Shadow (not shown in any of the photos). The Thermik is our flagship – it can carry a lot of equipment, including large batteries, and has a really sturdy wing. The Radian Pros are our test planes. They are very easy to operate and cheap to replace.
Please Microsoft, keep your blue screens away from our UAV hobby or businesses. We’ve been enjoying such a stable development so far
Is BRL suited to running at a much lower loop rate than the main loop rate? We’re thinking of adding a LowPriMath thread in ArduPilot that would soak up unused CPU cycles. Perhaps the BRL could go there?
Another option is to use one of the Linux based controller options for ArduPilot. Something like the NavIO on a RPi3 would make it easy to get lots of CPU but still have tight integration.
@tridge, BRL is suited for running at a lower rate than the main loop, but it requires quite it a bit of resources. Ultimately, the issue is that it scales roughly exponentially in the planning horizon, in terms of both CPU and memory. There are ways to reduce these requirements somewhat, especially for memory, but it is still too heavy to run on a Pixhawk comfortably.
Running ArduPilot on RPi3 is a more appealing option for BRL-based drone control – thank you, I’ll look into that! However, having LowPriMath would still be very useful! Lighter-weight approaches at the intersection of AI and control could make use of it.
Cool video, the Thermik looks like an awesome platform! Thanks again for the help getting AP_Soar into master. Looking forward to seeing the publications and maybe some more pull requests from your team in the future. Cheers, S
Agree. Code contributions back to ArduPlane would be welcome.