Slashdot Mirror
Intelligent Autonomous Flying Robots Learn and Map Environment As They Fly
An anonymous reader writes with this story about a machine-learning project out of the UK's University of Sheffield: Using simple drones the researchers have created automatic-control software that enables the "flying robot" to learn about its surroundings using a camera and an array of sensors. The robot starts with no information about its environment and the objects within it. But by overlaying different frames from the camera and selecting key reference points within the scene, it builds up a 3D map of the world around it. Other sensors pick up barometric and ultrasonic data, which give the robot additional clues about its environment. All this information is fed into autopilot software to allow the robot to navigate safely, but also to learn about the objects nearby and navigate to specific items.
It would be nice if these were fed back in real time to a remote monitor. Maybe a 21st century canary in a coal mine? Applications for search and rescue, scouting real time optimal traffic routes for police / fire / paramedics.
Cool project, but the article/video is short on detail. I'd like to know more about the way this robot is actually learning. Is it a neural network? How does it know an oscilloscope is an oscilloscope? Does it use binocular vision to recognize distance? Ultrasound? Both? What type of computing hardware is on board? For that matter, what type of quadracoptor is this? And more importantly where can I get one?
This and no other is the root from which a tyrant springs; when first he appears as a protector - Plato (423 to 327 BC)
Carnegie Mellon folks developed the SLAMM algorithm (and variants of if) some years back to do live mapping on their quadrotors. It has been used by almost everybody who is doing autonomous flying robots. It is hard not to say that anyone was not influenced by that work. Some of their work had laser scanners that would map the surroundings and identify walls -- building out a maze of sorts as it explored. Heck on seeedstudio.com you can pickup a (LIDAR) 360 2D laser scanner and algorithm to build your own.
Not taking anything away from the English (as what they have done is pretty impressive and a different approach), though am curious how it is different algorithm wise?
One of the research avenues I think at some point will be to have mesh networking on these robots to share info so that in a hostile/harsh environment as these things get damaged the mapping data is passed along for fire crews, soldiers, HAZMAT, etc.
Doing this is called Simultaneous Localization and Mapping, or SLAM. There's been enormous progress in that in the last decade. The basic idea is to take a large number of images of the same scene, possibly with inacccurate data about where they were taken, and build up a 3D model. It sort of works most of the time. Some algorithms do well indoors, especially where there are lots of strong edges and corners. Those are easy features to lock onto. Outdoors is tougher, although outdoors you can usually use GPS. It's a basic capabiilty robots need.
The video is frustrating. There's no comparison with previous work. Is this an advance, or did they just use known algorithms.
They tend to bump into the same walls repeatedly before learning they're there and proceeding to bump into the adjacent wall.
Corruption is convincing someone that the selfless ideal is the same as their selfish ideal.
So the robot starts with no information about its environment and the objects within it. By overlaying different frames from the camera and selecting key reference points within the scene, it builds up the 3D map of the world around it. Barometric and ultrasonic sensors give the robot additional clues about its environment. All this information is fed into autopilot software to allow the robot to navigate safely, but also to learn about the objects nearby and navigate to specific items.
Instead of a neural network, the researchers used just basic game theory to program the quadrocopters. In this framework, each robot is a player in the game and must complete its given task in order to "win" the game. If the robots "play the game" repeatedly they start to learn each other’s behavior. They can then perform their task successfully – in this case getting past the other robot – by using previous experience to estimate the behavior of the other robot.
As to your question about where to get one of these, this is just a research project. It would require a lot of more hard work, some Asian manufacturing arrangements, a global supply chain, and an investor to make this a product. Maybe some day.
Indeed, who would be interested in an vacuum cleaner capable of smoothly navigating in 3D space. At least design something practical, like a TV integrated on my fridge.
The more intelligent and autonomous my flying robots are, the better, I say ... hey, what are you ... Gahhhhhhhhh!!!
I want some of whatever it is you're smoking!
As soon as I saw the article I thought that it was just the sort of thing they had in Prometheus. It would be extremely useful for the military to be able to map out the inside of a building. Of course, you won't know what's behind closed doors, but they'll add something for that next.
They are using PTAM package from Uni of Oxford
http://www.robots.ox.ac.uk/~gk...
Whats more they are using off the shelf ardrone-PTAM package
https://github.com/nymanjens/a...
and replicating something done TWO YEARS AGO by Jens Nyman (from Belgian uni)
https://www.youtube.com/watch?...
so W T F
Who logs in to gdm? Not I, said the duck.
You want to get skynet? Cause this is how you get skynet.
Not too easily purposed to warfare and domination of other peoples. Just what we need more of.