Alex Chortos, assistant professor of mechanical engineering, develops robotic devices that interface with biological systems. “Soft robotics systems may have two or three actuators, but a real biological system like an octopus may have thousands of actuators,” he said. “We want to bridge that gap, so these robots can more accurately start mimicking biological systems. To do that, we need to be able to control many actuators at one time.”

His team’s experiment focuses on a flat surface that morphs into distinct 3D shapes with smooth waves, ridges, and indentations. In the future, such a device could function as a tactile feedback system, or even a control surface.  But because it is a smoothly graduated surface, it requires a lot more calculation than haptic systems today, which use simple on-or-off bumps.

Chortos began investigating this in 2020, during the COVID lockdowns. His Ph.D. student, Jue Wang, was unable to physically come to Purdue, and so he worked remotely on theoretical systems and simulation. Their research showed that machine learning can be trained to return the same results as traditional finite element analysis, but 15,000 times faster – fast enough to function as a real-time control system.

After the lockdowns were lifted, Wang was able to come to Purdue and build an experiment to verify the simulation. He fabricated a flexible sheet 5.4 centimeters square, divided into a six-by-six grid of pixels. Each of the 36 pixels was composed of an ionic electro-active polymer actuator, which flexed when a small voltage was applied. By sending voltage in different combinations, the surface could morph into selected shapes.

“This is very different from the way most soft robots are actuated today,” said Chortos. “If you had a linear actuator to cause something to pop up or down, you would need a large control system underneath it. With this system, the whole surface is an actuator, and it can be made paper-thin.”

Their research has been featured in ScienceAdvances.

“There are two innovations in this paper,” said Chortos. “First: how do we figure out how much power to send to these actuators to get them to morph into the shape we want? We developed the machine-learning algorithms to tackle that issue. But second: how do we physically deliver that power to the grid of 36 actuators? Would we need 36 tiny sets of wires coming out of this tiny square? That’s where passive matrix addressing comes in.”

Passive matrix addressing is a technology used by electronic ink displays like the Amazon Kindle. Rather than requiring a distinct value for each individual pixel, it addresses just the rows and columns. In other words, a six-by-six grid would only require 12 total voltage signals (rows + columns), rather than 36 (rows x columns). “This drastically reduces the number of wires going into the mechanism,” said Chortos. “Matrix addressing has been used in 2D optical displays, but it has not been used to simultaneously control arrays of physical actuators. As we increase the number of pixels for higher levels of detail, this will help to decrease the complexity of the device.”

“This only works because we chose actuators that stay actuated when we turn off the power,” said Wang. “Once we send voltages to the grid of pixels, they hold their shape – even after we disconnect the voltage. That’s a very important feature for this project.”

For the experiment, they duplicated their previous simulations where they attempted to make a letter “P” indented into the sheet. Using their combination of machine learning and passive matrix addressing, the physical six-by-six grid successfully deformed into a “P,” just like their simulations. It didn’t happen very fast – a few seconds per row of pixels – but for this particular experiment, the goal was just proving the concept behind the algorithms. The next step is integrating these theoretical control strategies into practical applications.

“We work on a range of actuator technologies and characteristics,” said Chortos. “Now we are looking to apply these concepts of machine learning and passive matrix addressing to a wide variety of applications. We can build all types of physical displays, wearable devices, and control surfaces. There really are no limits to this research.”

Wang is also looking forward to the next dimension. “We want to take what we developed for machine learning with this 2D structure, and build a 3D structure,” he said. “We have shown we can make a sheet deform into a ‘P’, so we hope to apply those same strategies to enable a cube to morph into any 3D object.”

Source: Alex Chortos, achortos@purdue.edu

Writer: Jared Pike, jaredpike@purdue.edu, 765-496-0374

Passively Addressed Robotic Morphing Surface (PARMS) Based on Machine Learning
Jue Wang, Michael Sotzing, Mina Lee, Alex Chortos
https://doi.org/10.1126/sciadv.adg8019
ABSTRACT: Reconfigurable morphing surfaces provide new opportunities for advanced human-machine interfaces and bio-inspired robotics. Morphing into arbitrary surfaces on demand requires a device with a sufficiently large number of actuators and an inverse control strategy. Developing compact, efficient control interfaces and algorithms is vital for broader adoption. In this work, we describe a passively addressed robotic morphing surface (PARMS) composed of matrix-arranged ionic actuators. To reduce the complexity of the physical control interface, we introduce passive matrix addressing. Matrix addressing allows the control of N² independent actuators using only 2N control inputs, which is significantly lower than traditional direct addressing (N² control inputs). Using machine learning with finite element simulations for training, our control algorithm enables real-time, high-precision forward and inverse control, allowing PARMS to dynamically morph into arbitrary achievable pre-defined surfaces on demand. These innovations may enable future implementation of PARMS in wearables, haptics,