I'm going to be a pedant here - that does not look like a swarm, and they don't claim it is (in the video, at least). By "swarm" I would understand that all have the same programming and that the assembly process is emergent from simpler instructions plus exploration and feedback. Instead they appear to be programmed to act sequentially with a pre-planned set of actions. There is no feedback, no emergence. In short, it's nowhere near as cool as you seem to think it is. It's just a bunch of dumb machines doing exactly what they are told to do, with no smarts.
No smarts? From their behavior, I got the impression that the robots were assessing what was the next step in construction, which included assessing if another robot was still working on construction. They also had to avoid collisions.
Anything programmed by humans is a "dumb machine doing exactly what they are told." I think you meant that to mean each individual step was hardcoded. I don't think it was. I think they were all following the same algorithm. The group certainly spends a lot of time on path planning: http://repository.upenn.edu/grasp_papers/
Correct about the swarm -- indeed it's not a multiagent problem -- but perhaps a bit unfair regarding the "dumb machine" part.
This appears to be a "multi-body" problem (rather than a "multi-agent") problem: that is, there is likely one primary decision-maker which dictates the construction task ordering, likely computed via a partial-order multi-body planner. That's part 1 of the AI.
But there's another part that's equally interesting. The multi-body planner produces a series of construction steps. Each step is ordered with regard to certain steps, unordered with regard to other steps (these can be done in any order but not simultaneously), and explicitly declared parallel with yet other steps (they can be done simultaneously). The quad-roters grab the steps and perform them in parallel when possible, and at least form a queue otherwise. Last, the quad-rotors have to perform these steps without hitting one another, so there's a significant real-time multi-body path-planning problem involved.
Multi-body planning, scheduling, and plan monitoring is nontrivial and should not be so lightly dismissed.
Hey, you can't have emergent properties before you have foundation properties. Can't build a building with magical emergent properties if you don't first figure out how to make a bot that can lift the beams.
This lab has several other videos with quad rotor robots like this (e.g. http://www.youtube.com/watch?v=MvRTALJp8DM), but don't worry about them escaping and starting a robot domination: they rely on infrared motion capture systems in that room for extremely precise location sensing.
Once GPS gets sub-meter accuracy, then we should start worrying.
Solutions based on expensive IMUs [0], SLAM [1], or visual odometry are more realistic than GPS. As a rule, never use GPS for fine-grained control. Even ignoring accuracy, GPS fails when you lack line-of-sight to GPS satellites, which is extremely dangerous on a quadrotor - how do you safely land a flying chainsaw with no feedback? When finely controlling a platform, you need accurate relative position. The global position is usually unimportant when the platform's position on Earth doesn't matter.
Sensors designed for computer vision applications are surprisingly light and small! Board-level cameras can fit in the palm of your hand, and reasonable IMUs are not much bigger than a matchbox. The biggest obstacles today are tradeoffs between performance/weight and performance/battery life. The platform needs algorithms with acceptable results, requiring almost no supporting hardware, and that are robust to the thousands of things that go wrong when trying to sense (noise, errors, resolution, occlusions, etc). Algorithms that fit all of these requirements are very uncommon. In fact, I would not be surprised if there was an Iron Triangle in there somewhere.
[0] Inertial Measurement Units: Devices that estimate acceleration (or some other measure of movement).
[1] Simultaneous Localization and Mapping. Using stereo vision, lidar, radio ranging, etc. to estimate your position, plus how the environment looks.
GPS + Accelerometer is already pretty darn accurate, I'm not really sure why they use the IR room. Perhaps simply easier to set up, and easier to co-ordinate the bots?
Pretty neat! I'm sure they'll get to it eventually, but there's no advantage to having three drones in this demo, since the work isn't pipelined. It seems that a drone has to wait until one finishes picking stuff up before it can pick up something itself.
Interesting, but robotic cranes on the ground would seem to be more efficient and reliable in ordinary circumstances. It's not a question of how to grip it, It's a simple question of weight ratios.
What you say makes sense - robotic cranes should be able to operate with heavier pieces, and will require less energy per piece placement.
However, MAVs (micro air vehicles) could reach into places where cranes wouldn't be able - i.e. inside the building.
Also you can have many more MAVs working at the same time than cranes, so MAVs might be able to considerably speed up the building process.
A couple of years ago I've watched a documentary about crane operators working in a harbors. They were payed incredible salaries and it was stated that they won't be replaced by robots because of their capabilities in decision making.
I've been cleaning up after a flood disaster for the past few days, the most amazing technology i've seen in that time is the bobcats and their drivers. It takes 10-15 people a few hours to empty a house of all it's waterlogged belongings on to the street, and then 15 minutes for the bobcat to get it up on the back of a truck. When I see a tech demo like this I'm very optimistic about it's potential for disaster relief, where simple structures, shelters, and platforms would be of great benefit.
Great fun, but I suspect NIMBY concerns would kill any practical application.
Let's imagine they get the machines scaled to the point when they could build an actual habitable structure, and that they sort the power concerns that I suspect would make it insurmountably uneconomic.
Now, imagine a squad of sufficiently large and powerful helicopters buzzing away all day next to your office. How many people are going to be OK with that?
Why couldn't small drones build things out of small elements? A lot of our assumptions about what is a reasonable unit of work are based on examples with humans in the loop. It might make more sense to think of a fleet of these things as a form of 3D printer.
So, human workers and ordinary, loud construction equipment buzzing away all day for 2 years is ok, but robot equipment - quite capable of working 24/7 and likely capable of going faster - buzzing away for 8 months, thats unacceptable? I think unions are gonna scream louder than any NIMBY concerns.
The shortsightedness of people worried about fewer jobs because of new technology is ridiculous. If you can build a house for the cost of land and materials only, the cost of living is lowered for everyone. With enough advancement and automation, the cost of life's necessities drops. As it decreases, the amount of human labor necessary to support day-to-day life will drop with it. Eventually we'll be able to support ourselves with very little laborious work, focusing instead on content production and other interesting and creative endeavors. Why try to prevent that?
Robotic construction will eventually take over the construction industry. Especially once robotic transportation becomes possible.
There are many complex processes being performed at construction sites. Many steps take highly optimized machines to perform robotically but that a single human can do just by changing tools.
Because of these processes, automating the entire construction process would be very expensive to do right at the site.
All parts would have be modularized so they could be snapped together. Or robots will have to be able to change tools. Or parts will have to be moved from robot to robot, with as much work being done before parts are actually brought to the site.
Either way, it will be interesting to see how it works out.
This is awesome! It's sort of like a less precise large-scale 3D printer or MineCraft. I'm sure I'm not the only person who thinks this is certainly the beginning of automated construction. It works for assembly lines. We may finally be getting to the point where the same process will work in more mobile volatile environments.
The power limitation doesn't seem very difficult to surmount, a drone could just go swap its dead battery with one from a pool of batteries on chargers, and keep working.
They'd need two batteries in a drone to do that of course, or perhaps some other robot could execute the battery swap.
In another video they show 4 of them working together to lift a 1.2 kg structure, and they don't seem to have much trouble with it. It also shows a single copter lifting a piece of 2x4 several inches long, but doesn't give the weight. As just a ballpark guess I'd say these things probably max out at lifting and flying with a little under half a kg, they don't seem to have any trouble with at least .25 kg. The columns in the video linked at the top of the page are most likey a lot lighter than that, simply because there's no need for them to be heavy.
[+] [-] andrewcooke|15 years ago|reply
[+] [-] scott_s|15 years ago|reply
Anything programmed by humans is a "dumb machine doing exactly what they are told." I think you meant that to mean each individual step was hardcoded. I don't think it was. I think they were all following the same algorithm. The group certainly spends a lot of time on path planning: http://repository.upenn.edu/grasp_papers/
[+] [-] SeanLuke|15 years ago|reply
This appears to be a "multi-body" problem (rather than a "multi-agent") problem: that is, there is likely one primary decision-maker which dictates the construction task ordering, likely computed via a partial-order multi-body planner. That's part 1 of the AI.
But there's another part that's equally interesting. The multi-body planner produces a series of construction steps. Each step is ordered with regard to certain steps, unordered with regard to other steps (these can be done in any order but not simultaneously), and explicitly declared parallel with yet other steps (they can be done simultaneously). The quad-roters grab the steps and perform them in parallel when possible, and at least form a queue otherwise. Last, the quad-rotors have to perform these steps without hitting one another, so there's a significant real-time multi-body path-planning problem involved.
Multi-body planning, scheduling, and plan monitoring is nontrivial and should not be so lightly dismissed.
[+] [-] sliverstorm|15 years ago|reply
[+] [-] yellowbkpk|15 years ago|reply
Once GPS gets sub-meter accuracy, then we should start worrying.
[+] [-] jakevoytko|15 years ago|reply
Sensors designed for computer vision applications are surprisingly light and small! Board-level cameras can fit in the palm of your hand, and reasonable IMUs are not much bigger than a matchbox. The biggest obstacles today are tradeoffs between performance/weight and performance/battery life. The platform needs algorithms with acceptable results, requiring almost no supporting hardware, and that are robust to the thousands of things that go wrong when trying to sense (noise, errors, resolution, occlusions, etc). Algorithms that fit all of these requirements are very uncommon. In fact, I would not be surprised if there was an Iron Triangle in there somewhere.
[0] Inertial Measurement Units: Devices that estimate acceleration (or some other measure of movement).
[1] Simultaneous Localization and Mapping. Using stereo vision, lidar, radio ranging, etc. to estimate your position, plus how the environment looks.
[+] [-] adamsmith|15 years ago|reply
You can get about 1cm of ranging error (median), and say 10cm of positioning error.
[+] [-] sliverstorm|15 years ago|reply
[+] [-] asdkl234890|15 years ago|reply
Or once the kinect shrinks to size of a short stack of quarters.
[+] [-] iamwil|15 years ago|reply
[+] [-] AdamTReineke|15 years ago|reply
[+] [-] mhansen|15 years ago|reply
[+] [-] brudgers|15 years ago|reply
[+] [-] mitko|15 years ago|reply
However, MAVs (micro air vehicles) could reach into places where cranes wouldn't be able - i.e. inside the building. Also you can have many more MAVs working at the same time than cranes, so MAVs might be able to considerably speed up the building process.
[+] [-] T-zex|15 years ago|reply
[+] [-] bpd1069|15 years ago|reply
Once you are free of that limitation, applications of this technique on varying scales opens up a new "world" of possibilities.
Pipelining/Parallelization could be achieved by throwing more bots at the problem.
[+] [-] stcredzero|15 years ago|reply
[+] [-] d5tryr|15 years ago|reply
[+] [-] netcan|15 years ago|reply
[+] [-] thebigredjay|15 years ago|reply
[+] [-] eftpotrm|15 years ago|reply
Let's imagine they get the machines scaled to the point when they could build an actual habitable structure, and that they sort the power concerns that I suspect would make it insurmountably uneconomic.
Now, imagine a squad of sufficiently large and powerful helicopters buzzing away all day next to your office. How many people are going to be OK with that?
[+] [-] pg|15 years ago|reply
[+] [-] philwelch|15 years ago|reply
[+] [-] jpitz|15 years ago|reply
[+] [-] jaekwon|15 years ago|reply
[+] [-] siculars|15 years ago|reply
[+] [-] BrandonM|15 years ago|reply
[+] [-] ph0rque|15 years ago|reply
[+] [-] onteria|15 years ago|reply
http://www.grasp.upenn.edu/research
[+] [-] btipling|15 years ago|reply
[+] [-] stevenbedrick|15 years ago|reply
[+] [-] stretchwithme|15 years ago|reply
There are many complex processes being performed at construction sites. Many steps take highly optimized machines to perform robotically but that a single human can do just by changing tools.
Because of these processes, automating the entire construction process would be very expensive to do right at the site.
All parts would have be modularized so they could be snapped together. Or robots will have to be able to change tools. Or parts will have to be moved from robot to robot, with as much work being done before parts are actually brought to the site.
Either way, it will be interesting to see how it works out.
[+] [-] wglb|15 years ago|reply
[+] [-] shaunfs|15 years ago|reply
[+] [-] soamv|15 years ago|reply
They'd need two batteries in a drone to do that of course, or perhaps some other robot could execute the battery swap.
[+] [-] fara|15 years ago|reply
[+] [-] johnohara|15 years ago|reply
[+] [-] burgerbrain|15 years ago|reply
EDIT: that piece of 2x4 looks to be about 12 inches long. Assuming the wood is pine (likely, and also light), and using 350kg/m^3 as the pine's density (http://www.simetric.co.uk/si_wood.htm the lightest figure quoted) that works out to about .55 kg. (http://www.wolframalpha.com/input/?i=350+kg%2Fm^3+*+%282*4*1...). Quite impressive.
[+] [-] hanula|15 years ago|reply
[+] [-] fakespastic|15 years ago|reply