Movie fanatics still remember how the two AI humanoid robots – TARS and CASE  – assisted Joseph Cooper (Matthew McCauneghey) in Interstellar. The two characters were built by the team as real machines operated by a puppeteer, erasing the human operator with CGI (computer-generated imagery), and adding functions that couldn’t be achieved in real life. Christopher Nolan and members of the crew chose the ‘minimalist approach’ to design the military robots. Explaining the philosophy behind the two robots, Nolan says, “To me, the idea of a robot is a machine impersonating a human being. These are purely machines.”

The same is the case with manufacturing satellites and spaceships that involve multiple repetitive operations with high precision. Moreover, most of the process must be performed in isolated rooms, in order to avoid potential contamination by microscopic life forms. 

According to NASA, the biological pollution of cosmic bodies such as the Moon or Mars can distort research data on alien life. In order to minimize contact with people, companies have started using systems based on AI algorithms while assembling vehicles like spacecraft. At present, neural networks form the basis of many modern AI set-ups. The same concept has now been applied to a mechanical calculating machine, bringing the foundations of AI in mechanics one step closer to learning behaviors without human aid. 

An ‘AI Muscle Memory’ 

Artificial neural network (ANN) is one of  the basics of modern AI research which creates large grids of artificial neurons to mimic the structure of the human brain. Similar to the brain which learns new behaviors by reinforcing synaptic connections, ANNs learn by adjusting the stored digital values to represent them. 

Imagine a plane wing that catches a gust of wind and then is forced to move in an unanticipated direction. It’s not possible for the wing to change its design to be stronger. Hypothetically, a new prototype lattice material can keep the plane from changing to unknown conditions. Through successive adjustments made by the algorithm periodically, the wing adopts new properties, by adding each behavior as a sort of muscle memory.

A group of researchers from the University of California and the University of Twente, in their paper ‘Mechanical neural networks: Architected materials that learn behaviors’, introduced an adaptive material — ‘mechanical neural network’ – that can actively respond to changing conditions. Equivalent to its software capabilities, the mechanical neural network is composed of beams and sensors capable of learning to carry out several different tasks. According to the paper, this new innovation could lead to aircraft wings that morph to maintain efficiency and minimize turbulence. Moreover, the new adaptive material in AI is capable of learning this on its own.

Researcher Ryan Lee at University of California, Los Angeles, said that the team borrowed the concept from ANNs to create a neural network where beams of variable stiffness are the strength of connections between each neuron. Rather than processing digital data, the mechanical neural network processes the force applied – depending on the stiffness of its beams – to twist and morph from its original shape. 

Aircraft wing with morphable capabilities

Lee says that an aircraft wing manufactured on this system could automatically morph in response to situations, by preventing undesirable flight characteristics or boosting efficiency. “You could do something neat like turbulence interference, where the wing gets hit by something and locally deforms and morphs to try and keep the energy spread out in a way that the cabin feels nice and smooth. Right now, wings are designed to do big motions, flex, distribute that across the wing, and that results in jerkiness in the cabin.”

The network is made of 21 beams (each 15 centimeters long) arranged in a triangular grid or lattice. Every individual beam is equipped with a small linear motor capable of altering its stiffness, along with sensors that measure how each “neuron” is out of position. Thus, the computer is able to train the network by tweaking the beam stiffness. Once done, the structure needs no external computation as the beam stiffnesses are locked in.

Source: Science Robotics

Research lead and mechanical and aerospace engineering professor at the UCLA Samueli School of Engineering, Jonathan Hopkins, said, “This research introduces and demonstrates an artificial intelligent material that can learn to exhibit the desired behaviors and properties upon increased exposure to ambient conditions. The same foundational principles used in ML are implemented to give this material its smart and adaptive properties.”

Source: Flexible Research Group, UCLA

The team has spent five years ironing out the idea of how the mechanical neural network material can learn or react in real-time. 

3D vision seems far-sighted

The new algorithm model is all set to scale up. Perhaps, like every other model, lies an exception – until now, the team has worked only with 2D lattices. Lee says, “But using computer modeling, we predict that 3D lattices would have a much larger capacity for learning and adaptation. This increase is due to the fact that a 3D structure could have tens of times more connections that don’t intersect with one another. However, the mechanisms we used in our first model are far too complex to support in a large 3D structure.”

However, the team believes that the material has created a proof of concept, highlighting the potential of MNNs. Bringing this idea into reality will require generating smaller individual pieces with precise properties of flex and tension. He hopes that manufacturing of materials at the micron scale will lead to powerful smart MNNs with dense 3D connections – a possible reality in the near future.