Science is exciting in theory, but it can also be dreadfully dull. Some experiments require hundreds or thousands of repetitions or trials — an excellent opportunity to automate. That’s just what MIT scientists have done, creating a robot that performs a certain experiment, observes the results, and plans a follow-up… and has now done so 100,000 times in the year it’s been operating.
The field of fluid dynamics involves a lot of complex and unpredictable forces, and sometimes the best way to understand them is to repeat things over and over until patterns emerge. (Well, it’s a little more complex than that, but this is neither the time nor the place to delve into the general mysteries of fluid dynamics.)
One of the observations that needs to be performed is of “vortex-induced vibration,” a kind of disturbance that matters a lot to designing ships that travel through water efficiently. It involves close observation of an object moving through water… over, and over, and over.
Turns out it’s also a perfect duty for a robot to take over. But the Intelligent Tow Tank, as they call this robotic experimentation platform, is designed not just to do the mechanical work of dragging something through the water, but to intelligently observe the results, change the setup accordingly to pursue further information, and continue doing that until it has something worth reporting.
“The ITT has already conducted about 100,000 experiments, essentially completing the equivalent of all of a Ph.D. student’s experiments every 2 weeks,” say the researchers in their paper, published today in Science Robotics.
The hard part, of course, was not designing the robot (though that was undoubtedly difficult as well) but the logic that lets it understand, at a surface level so to speak, the currents and flows of the fluid system and conduct follow-up experiments that produce useful results.
Normally a human (probably a grad student) would have to observe every trial — the parameters of which may be essentially random — and decide how to move forward. But this is rote work — not the kind of thing an ambitious researcher would like to spend their time doing.
So it’s a blessing that this robot, and others like it, could soon take over the grunt work while humans focus on high-level concepts and ideas. The paper notes other robots at CMU and elsewhere that have demonstrated how automation of such work could proceed.
“This constitutes a potential paradigm shift in conducting experimental research, where robots, computers, and humans collaborate to accelerate discovery and to search expeditiously and effectively large parametric spaces that are impracticable with the traditional approach,” the team writes.
If you want to build a robot that can fall hundreds of feet and be no worse the wear, legs are pretty much out of the question. The obvious answer, then, is a complex web of cable-actuated rods. Obvious to Squishy Robotics, anyway, whose robots look delicate but are in fact among the most durable out there.
The startup has been operating more or less in stealth mode, emerging publicly today onstage at our Robotics + AI Sessions event in Berkeley, Calif. It began, co-founder and CEO Alice Agogino told me, as a project connected to NASA Ames a few years back.
“The original idea was to have a robot that could be dropped from a spacecraft and survive the fall,” said Agogino. “But I could tell this tech had earthly applications.”
Her reason for thinking so was learning that first responders were losing their lives due to poor situational awareness in areas they were being deployed. It’s hard to tell without actually being right there that a toxic gas is lying close to the ground, or that there is a downed electrical line hidden under a fallen tree, and so on.
Robots are well-suited to this type of reconnaissance, but it’s a bit of a Catch-22: You have to get close to deploy a robot, but you need the robot there to get close enough in the first place. Unless, of course, you can somehow deploy the robot from the air. This is already done, but it’s rather clumsy: picture a wheeled bot floating down under a parachute, missing its mark by a hundred feet due to high winds or getting tangled in its own cords.
“We interviewed a number of first responders,” said Agogino. “They told us they want us to deploy ground sensors before they get there, to know what they’re getting into; then when they get there they want something to walk in front of them.”
Squishy’s solution can’t quite be dropped from orbit, as the original plan was for exploring Saturn’s moon Titan, but they can fall from 600 feet, and likely much more than that, and function perfectly well afterwards. It’s all because of the unique “tensegrity structure,” which looks like a game of pick-up-sticks crossed with cat’s cradle. (Only use the freshest references for you, reader.)
If it looks familiar, you’re probably thinking of the structures famously studied by Buckminster Fuller, and they’re related but quite different. This one had to be engineered not just to withstand great force from dropping, but to shift in such a way that it can walk or crawl along the ground and even climb low obstacles. That’s a nontrivial shift away from the buckyball and other geodesic types.
“We looked at lots of different tensegrity structures — there are an infinite number,” Agogino said. “It has six compressive elements, which are the bars, and 24 other elements, which are the cables or wires. But they could be shot out of a cannon and still protect the payload. And they’re so compliant, you could throw them at children, basically.” (That’s not the mission, obviously. But there are in fact children’s toys with tensegrity-type designs.)
Inside the bars are wires that can be pulled or slackened to cause to move the various points of contact with the ground, changing the center of gravity and causing the robot to roll or spin in the desired direction. A big part of the engineering work was making the tiny motors to control the cables, and then essentially inventing a method of locomotion for this strange shape.
“On the one hand it’s a relatively simple structure, but it’s complicated to control,” said Agogino. “To get from A to B there are any number of solutions, so you can just play around — we even had kids do it. But to do it quickly and accurately, we used machine learning and AI techniques to come up with an optimum technique. First we just created lots of motions and observed them. And from those we found patterns, different gaits. For instance if it has to squeeze between rocks, it has to change its shape to be able to do that.”
The mobile version would be semi-autonomous, meaning it would be controlled more or less directly but figure out on its own the best way to accomplish “go forward” or “go around this wall.” The payload can be customized to have various sensors and cameras, depending on the needs of the client — one being deployed at a chemical spill needs a different loadout than one dropping into a radioactive area, for instance.
To be clear, these things aren’t going to win in an all-out race against a Spot or a wheeled robot on unbroken pavement. But for one thing, those are built specifically for certain environments and there’s room for more all-purpose, adaptable types. And for another, neither one of those can be dropped from a helicopter and survive. In fact, almost no robots at all can.
“No one can do what we do,” Agogino preened. At a recent industry demo day where robot makers showed off air-drop models, “we were the only vendor that was able to do a successful drop.”
And although the tests only went up to a few hundred feet, there’s no reason that Squishy’s bots shouldn’t be able to be dropped from 1,000, or for that matter 50,000 feet up. They hit terminal velocity after a relatively short distance, meaning they’re hitting the ground as hard as they ever will, and working just fine afterwards. That has plenty of parties interested in what Squishy is selling.
The company is still extremely small and has very little funding: mainly a $ 500,000 grant from NASA and $ 225,000 from the National Science Foundation’s SBIR fund. But they’re also working from UC Berkeley’s Skydeck accelerator, which has already put them in touch with a variety of resources and entrepreneurs, and the upcoming May 14 demo day will put their unique robotics in front of hundreds of VCs eager to back the latest academic spin-offs.
You can keep up with the latest from the company at its website, or of course this one.
When humanity’s back is against the wall and the robots have us cornered I’d say I’m all for whanging a few with a baseball bat. However, until then, we must be kind to our mechanical brethren and this robotic tortoise will help our kids learn that robot abuse is a bad idea.
Researchers at Naver Labs, KAIST, and Seoul National University created this robot to show kids the consequences of their actions when it comes to robots. Called Shelly, the robot reacts to touches and smacks. When it gets scared it changes color and retracts into its shell. Children learn that if they hit Shelly she will be upset and the only thing missing is a set of bitey jaws.
“When Shelly stops its interaction due to a child’s abusive behavior, the others in the group who wanted to keep playing with Shelly often complained about it, eventually restraining each other’s abusive behavior,” Naver Labs’ Jason J. Choi told IEEE. The study found that Shelly’s reactions reduced the amount of abuse the robot took from angry toddlers.
The researchers showed off Shelly at the ACM/IEEE International Conference on Human Robot Interaction last week.
Science can be cute as hell when it wants to be – take the JEM Internal Ball Camera (“Int-Ball” for short). The device, created by the Japan Aerospace Exploration Agency (JAXA), was delivered to the International Space Station on June 4, 2017, and now JAXA is releasing its first video and images. The purpose of Int-Ball is to give scientists on the ground the ability to… Read More
The Vinobot rover teams up with an eye in the sky to keep tabs on corn so that one day humanity will still even have corn. The post The Busy Little Robot on a Quest to Help Humanity Feed Itself appeared first on WIRED.