Soft robotics – machines made of flexible, muscle-like materials – can bend and stretch in fluid ways that put the rigid robots of old sci-fi movies to shame.
But the flexibility that lets them pick ripe tomatoes or navigate a search-and-rescue site comes at a cost: soft robotics are notoriously difficult to control.
Virginia Tech researchers are working to solve that problem using a new computing approach inspired by the complex architecture of neurons in the brain.
Called reservoir computing, the technique allowed a team based in the Department of Mechanical Engineering to create a simulated robotic arm that can flex, twist, warp, and bend.
Not only did reservoir computing blow past the limits of conventional artificial intelligence (AI) and machine learning methods, when the team implemented the reservoir on a neuromorphic computer chip that spikes like the brain, it reduced power use by up to 75 times.
The research, which was published in the Proceedings of the National Academy of Sciences, could pave the way for the development of more small, untethered robots with applications in fields like medicine, agriculture, salvage, and infrastructure inspection.
“We don’t know if our approach is the best, but it is the first to be able to control this type of very flexible, very quickly moving soft arm,” said Noel Naughton, assistant professor of mechanical engineering, who led the research.
A different breed of robot
Soft robotics are built differently from the blocky, metallic robots of popular imagination. Made with a combination of soft materials and novel controls, soft robotics have a greater range of motion and more fluidity and flexibility than traditional rigid robotics.
They can deform and reshape, wrapping around objects instead of clamping down, which makes them an asset in areas that are dangerous or physically impossible for humans to reach.
The issue with these novel robotics is control. In traditional robotics, the movement of an arm or a finger is built on commands: Raising an arm might be one command or a series of them. With soft robotics, all that flexibility requires more complex controls.
Previously, Naughton used virtual tools and motion mapping to engineer new kinds of robotics. In the past, he has taken cues from an octopus to design robots that move in a similar way.
This time, Naughton’s team used those 3D virtual tools to build a simulated arm modeled after the anatomy of animals such as snakes.
Their arm uses a central elastic core with multiple pairs of synthetic muscles, similar to biceps and triceps in humans, that overlap and work together to move the arm.
The team’s goal was to determine the best method for automated, dynamic control of the arm; they wanted a way to contract and relax the simulated muscles around the core to make the arm twist and bend.
“When we put this idea together, we realized there was no known way to control it,” Naughton said.
A computer inspired by the brain
That required Naughton to take a fresh approach to the control issue: a neural reservoir.
In a neural reservoir, researchers input data about the movement of virtual soft robots, set parameters for what they expected to happen, ran virtual trials, and then analyzed the results.
Members of Naughton’s team knew the properties of the elastic core and the synthetic muscles as well as how those materials responded to bending and twisting. But they didn’t know the dynamics of how their muscle pairs would work together.
Using the neural reservoir, they created virtual models of different variations of movement and tested how they behaved.
When they fed those results back into the system, a new model for the behavior of a soft robotic arm began to emerge – along with a new approach to the most effective way to control the arm.
Neural computing is faster than building tomes of commands, and it’s also more energy efficient, using dramatically less power than traditional computers.
While for now the muscle-bound robotic arm is virtual, the data that Naughton’s team built will eventually be used to make physical robotics move.
“Now that we have these new tools, the next step is to build physical prototypes to test out our reservoir control approach on soft robotic arms,” Naughton said.
“Hopefully this will help close the gap between current soft robots and the amazing dexterity we see from soft biological creatures such as the octopus.”
