Slashdot Mirror
DARPA-Funded Software Could Usher In the Era of Open-Source Robotics
malachiorion writes "The best thing to come out of the DARPA Robotics Challenge, so far, isn't the lineup of nifty rescue bots being developed by teams around the world, or even Boston Dynamics' incredible Atlas humanoid. It's the pumped-up version of Gazebo, the free, open-source robotics simulation software whose expansion and further development is being funded by DARPA. This article has a look at how the software was used in the recent virtual leg of the competition, as well as how it could change the way robotics R&D is conducted (and create more roboticists, with its low-cost, cloud-based architecture)."
The Boeing 707 is a perfect example. The military demand for jets like the B-52 and KC-135 projects heavily influenced the iconic jetliner, and today, the 737 retains many of its design features. The 707 design is so similar to the KC-135 that upon their replacement with newer airliners, most examples were purchased by the military and used as parts for the KC-135 fleet.
Meanwhile, the Internet exists because of DARPA, the Blue Riband is held by a ship designed specifically for conversion to a troopship, and we enjoy any number of advances in a plethora of fields as a result of the space program, originally having its roots not only in the German rockets that attacked Britain but in showing the Russians how far we could lob nuclear warheads with the exact same missile designs.
I'm glad this time it's going open source.
In SOVIET RUSSIA... erm...NSA AMERICA, the Internet logs onto YOU!
Because if using the internet has taught us anything it's that when Skynet becomes self aware it's going to need to share its changes with the community.
i for one, welcome our new opensource robotic overlords with backdoor NoStringsAttached access
What DARPA is providing for Gazebo is adult supervision. They're paying for getting the bugs fixed. Gazebo has been around for years, and like most open-source projects with modest user bases, it sort of worked. Now it's finally getting fixed. It still only installs easily on Ubuntu 12.04, has tons of dependencies including limitations on supported graphics cards, has lots of bugs, and way too many configuration files. But it's now usable.
The "cloud" business is merely a way to make the DARPA competition honest. For competition purposes, the simulator runs in an Amazon AWS instance controlled by DARPA, with the simulated robot controlled through an API that only provides information a real robot would provide. The robot control programs written by competitors can't see the map of the world; all it gets is simulated vision and LIDAR data. It's a lot like the server/client relationship of an MMORPG. Each user has their own server instance; the world is not, as yet, shared.
The "cloud" is not otherwise necessary, or even desirable. For development purposes, you'd usually run the simulator and the control programs on the same machine, or at least a local machine.
A big problem with Gazebo is that the physics engine is only game-quality. Here it matters, because foot/ground contact is what supports the simulated robot, and most game simulators don't do contacts very well. Gazebo is in the process of switching from ODE to Bullet, which should help.
http://www.gazebosim.org/
http://en.wikipedia.org/wiki/Robotics_simulator
The real time simulation of an Atlas walking robot, with its walking controller, Laser scanner and 6 cameras (2 in each hand!) is beyond the capabilities of most ordinary computers. In fact, even the AWS instances were too slow and the contest was actually held on the SoltLayer infrastructure (now IBM), using K10 Tesla gpus and dual Intel Xeon E5-2690-OctoCore [2.9GHz]. The cloud is a good place is good place for simulation, not only because you can get a hundred computers for 1 hour (instead of having to wait 100 hours on a single one), but also because of the vast data sets you will eventually be able to tap into (grasping libraries, objects recognition data sets,etc). There's only a very finite amount of data you can upload and download to your local machine. Running robot simulations in the cloud is a great way to share. Publishing simulations and algorithms so that anyone can try them on a public cloud is much better than trying to get others replicate your local setup with tutorials and instructions. Not to mention the time you don't have to spend installing and configuring software.
yall still got me.
-dirtbag
This raises a related question - are there any pre-composed open source licenses out there that expressly prohibit any military use? I think its great and all that you kids are getting your money for playing around and doing what you enjoy, but at the end of the day have you considered the full consequences of what you're working toward?
Normally I keep to side lines and let the internet buzz take care of itself. However, I was disheartened to see the most highly ranked comment be so negative about a free project designed to improve not only robotics research, but also disaster response robots.
Animats does have one valid point. Gazebo has been buggy, in the past. It was also just a side project with one developer and no funding until a few years ago when Willow Garage brought Gazebo under its roof. Gazebo has now matured beyond a research project into a robust application with a growing community of users and developers.
It's been particularly exciting to have simulation move into the cloud. The cost and time required to start a complex simulation has been greatly reduced. Educators, hobbyists, and researchers now have a great tool for a wide variety of uses. Cloud computing is designed for large data sets and strict performance requirements. Simulation needs both of these.
ODE is currently the primary physics engine in Gazebo, with Bullet integration almost complete and SimBody and RTQL8 within a year. It deserves to be mentioned that ODE and Bullet both use a maximal coordinate solvers. This means they effectively produce the same result. Many games use engines similar or identical to ODE and Bullet. The key difference is how they are used. Gazebo optimizes ODE and Bullet for robot simulation. This means physics run slower, but we get out less residual error. There are other tricks game engines play that a simulator can't, but at the end of the day we are all just solving the equations of motion.
Over the next few years expect to see great things from Gazebo, including scripting interfaces, plotting utilities, robot modeling widgets, physics auto-tuning, many more robots and environments, vehicle suspension models, improved friction modeling, a sensor noise model pipeline, and deformable objects.
Most importantly, I want to thank everyone who have used and contributed to Gazebo (especially all the VRC teams who toiled through countless hours to produce very impressive results of Atlas completing complex tasks). Good luck to everyone continuing on in the DRC!!
You are really narrow-sighted about the concept of cloud simulation. Yes, it's a tool to host the VRC. It also has many other benefits for education at the highschool and college level, and hosting other educational competitions. With a little spit and polish it could be a be a mechanism to crowd-source robot design, and environment building. I haven't even touch upon the concept of running many simulation simultaneously as a prediction tool for physical robots.
I would love a cloud based simulation solution for development purposes. Please, spare me from the tedium of install and maintaining software. All I want to do is get work done.
Just because it can run on Amazon AWS doesn't mean it's free. Each job running in simulation ties up three large AWS machines and one small one. You'd have to rent machine time from Amazon and load up your own instances. There's no free world simulator to connect to. This thing takes a lot of engine behind it.