A screen shot of a humanoid robot turning a valve in the Virtual Robotics Challenge.
A simulator intended to give robot makers a hand during a Defense Advanced Research Projects Agency contest is expected to remain an integral part of how developers design the machines in the future.
DARPA is in the midst of its Robotics Challenge program, which is asking teams to use a humanoid robot to perform a series of complex disaster-response tasks.
There will be two competitions, one in December 2013 and a second in December 2014, where teams will take bipedal robots and put them through an obstacle course.
DARPA wanted to include as many entrants as possible, said Brian Gerkey, the CEO of the Open Source Robotics Foundation, a San Francisco Bay area nonprofit devoted to standardizing software in the industry.
The problem was that few companies, universities or nonprofits that wanted to participate had the acumen or the funds to build such a complex machine.
“Historically, for a team to compete in a challenge like this, they have to have the ability to do everything themselves,” he said. They would have to design and build the hardware, create the software, have an operations team to run experiments and have all the funding in place to do all these tasks.
There are teams that are really good at just the software side of robot development, he said.
The Robotics Challenge is similar to DARPA’s Grand Challenges, a series of three driverless car demonstrations that took place during the previous decade. In that competition, autonomous cars drove on closed courses.
The Robotics Challenge will ask teams to use a human-shaped robot to perform a series of disaster response tasks. In the original challenge, no teams were expected to actually build a vehicle from the ground up. They used commercially available cars or SUVs. Similarly, the organizers didn’t expect all the teams to have humanoid robots, Gerkey said.
“There are very few people in the world who know how to build walking humanoid robots. But there are probably many more people who are good at writing software for them,” he added.
DARPA asked the foundation to adapt training and simulation software in order to lower the barrier of entry for the contest.
There were four categories of participants. The first were those who were building the humanoid robots with funding from DARPA. The second were those who were creating software with funding from DARPA. The third wrote software without funding. The fourth are those doing it all — the software and the robots — with their own money.
The foundation created and organized the Virtual Robotic Challenge for those in the second and third categories, who had software but didn’t have the means to test it on a real humanoid robot.
It chose the Gazebo 3-D robotics simulation software as a basis and set up an exercise based on the 2011 Fukushima Daiichi Nuclear Power Plant disaster.
Japan is renowned for its robotics prowess. YouTube is full of videos from the country showing humanoid robots dancing, walking up stairs and playing musical instruments. But after the earthquake and tsunami struck, the nation had no disaster response robots it could send into the radioactive areas at the plant to shut down crucial valves.
If there had been robots that could have been dispatched and controlled remotely with experts on the other end, they could have done relatively simple tasks such as open doors and open and close valves, Gerkey said.
Eventually, U.S.-built robots normally used to dismantle bombs were sent in. They arrived late, though.
The nuclear disaster that followed the earthquake and tsunami could have been avoided if the right kind of robots were on the scene within the first few hours, Gerkey asserted.
“It was the lack of that technology that led to the meltdown that occurred,” he said. “Fundamentally, what I want to do is action at a distance. But I need a proxy for myself to be onsite doing the work.”
The reason for the humanoid robot rather than the squat wheeled or tracked machines associated with bomb disposal teams is that humans created the environments in which they would be employed. The manmade world — even when degraded after a disaster — tends to be scaled for humans, and is full of objects designed for people to manipulate. A disaster response robot would therefore look more like a human with some arms, some hands, and “maybe more arms, more hands, who knows?” Gerkey said.
The foundation set up the virtual portion of the competition in June.
Teams with software they wanted to test were given a series of tasks they had to perform with the virtual humanoid robot. They also were given some time to practice. They could only see on the 3-D viewer what the robot saw with its sensors. To further level the playing field, it was all done on cloud-based servers because not all the teams had the computing power or graphics cards to run the simulation.
They had 15 tasks over the course of three days such as walking over uneven terrain, getting into a vehicle and driving it and manipulating a fire hose.
They didn’t know which tasks would be in what order until they began.
There were more than 100 teams participating, and most of them had no funding from DARPA.
“If you only let in the teams that got funding from DARPA, then you’ve selected groups that know how to write effective DARPA proposals, which is not the same necessarily as being the best at writing robot software,” Gerkey said.
A preliminary phase whittled the number down to 26. A second round of tests resulted in DARPA choosing seven winners. Of those, two of the teams were unfunded.
The top six winners earned DARPA funding and an Atlas humanoid robot built by Boston Dynamics to use in the first round of real-world tests to take place Dec. 20-21 in Miami. The seventh-, eighth- and ninth-place winners also acquired Atlas robots through other means such as donations. They will be put through an obstacle course similar to the simulation they completed in June, Gerkey said. The nine Atlas robots will compete with the DARPA-funded humanoid robots built from the ground up and teams that are coming with their privately funded machines.
Kevin Knoedler, co-lead for the second-place Worcester Polytechnic Institute-Carnegie Mellon team, said his researchers would have had to create a simulator to test its software, which would have eaten up a lot of their time.
“The simulator was not perfect — they never are — but it allowed us to spend more time focusing on the tasks rather than doing simulator development,” he said.
As a top-tier winner, the team has received an Atlas robot. Yet it continues to use the simulator as the December competition approaches.
Running experiments on the Atlas takes time to set up. These tests can also break the hardware, he added. The simulation is used on a more than daily basis to bypass the real-world tests, he said. The teams’ researchers can also develop tasks in parallel, which creates further efficiencies, he noted.
“Everyone has a robot on [his] desktop now,” he added.
Results still have to be validated on the physical robot, he added, but the simulator has been invaluable in preparing for the December challenge, he said.
Gerkey said this could be one of the lasting legacies of the Robotics Challenge. The open source software will remain for researchers to use, and the foundation will continue to fine tune it.
“We are looking long term so it will serve as broad a group of people as possible in the robotics industry,” Gerkey said.
Simulators used to validate electronic circuits have revolutionized that industry, he said.
“Everyone does everything in simulation software and it is only when they are 100 percent sure that it works that they actually produce a circuit board. This has led to a massive increase in productivity in circuit design,” he said.
“If we can do the same thing in robotics, the cheapest robot out there is the one that exists in your software on your computer,” Gerkey said.
Knoedler added: “It will allow more people to do more work in the robotics field — humanoid robotics, wheeled robotics — all sorts of different types.”
Photo Credit: DARPA