The Pentagon Helped Fund Research That May Have Solved Soft Robotics' Biggest Problem

The Office of Naval Research helped pay for it. The university press release does not lead with that.

A team at Virginia Tech and the University of Illinois Urbana-Champaign published research in May 2026 in the Proceedings of the National Academy of Sciences describing a soft robotic arm that neural reservoir computing (an older AI method that exploits the robot's own physical dynamics as part of the computation rather than treating them as obstacles) could control 75 times more efficiently than the best conventional neural network approach. The paper lists joint funding from ONR MURI grant N00014-19-1-2373 and ONR grant N00014-22-1-2569 (alongside NSF grants) in its acknowledgments, not its introduction. Military interest in soft robots that can navigate confined underwater spaces and operate where rigid machines cannot is documented in ONR's own program literature. A small, efficient, untethered soft robot built for those environments has obvious applications beyond civilian inspection.

The distinction matters because of how each approach handles energy. Training a conventional neural network requires repeated passes over the same data, with the model adjusting millions of parameters through a process called backpropagation. That process is computationally expensive and power-hungry. Reservoir computing sidesteps it entirely: only the final output layer is learned. The reservoir itself (a recurrent neural network with fixed, untrained weights) does the work of transforming inputs into useful signals. Multiple independent control policies can run on the same fixed reservoir simultaneously.

"Soft robots are hard to control because they deform continuously as they move," said Noel Naughton, now at Virginia Tech's Department of Mechanical Engineering and the paper's lead author, whose research was conducted at the UIUC Beckman Institute. "The shape changes everything, the physics, the geometry, the forces. A conventional neural network trying to learn motor control has to account for that deformation on the fly. A reservoir computer doesn't. It exploits the robot's own physical dynamics as part of the computation."

When the team ran the same reservoir on Intel Loihi neuromorphic hardware (a chip that processes information the way a brain does rather than like a conventional processor), power consumption dropped by up to 75 times compared to standard CPUs and about 45 times compared to high-efficiency conventional processors. The paper describes a simulated soft arm built from 16 muscle-tendon units wrapped around an elastic spine; the team has not yet tested the reservoir controller on a physical robot outside the lab. The next step is building and testing physical prototypes.

For field robotics, the implications are straightforward on paper. Soft robots made from compliant materials can squeeze through gaps rigid machines cannot, adapt their shape to uneven terrain, and handle contact with humans without the injury risk of metal arms. Those capabilities have made them attractive for applications in medicine, agriculture, salvage operations, and infrastructure inspection. The problem has been that useful soft robots tend to be tethered to external power and computing. A robot that can only operate near a wall outlet has limited use inside a collapsed building or inside a human body.

Untethered operation requires solving motor control with onboard computing, and onboard computing requires either a large battery or an efficient processor. The new work suggests neuromorphic hardware running reservoir-based control is a plausible path to the latter. If the efficiency numbers hold on real hardware, not just simulations, the combination could enable small, battery-powered soft robots that deploy where tethering a cable is not an option.

Whether the 75-times efficiency number survives contact with a real robot in a real environment remains an open question. The Intel Loihi chip used in the study is research hardware, not a commercially deployed neuromorphic product. Scaling from a simulated arm to a physical robot introduces complications, sensor noise, material fatigue, actuator lag, that the simulation does not fully capture. And the baseline the researchers compare against matters: "standard CPUs" is a broad category, and a more carefully optimized conventional controller might narrow the gap.

But the directional finding is difficult to dismiss. For a class of problems where the robot's own physical dynamics are part of the computation, an approach that exploits those dynamics directly is more efficient than one that treats them as obstacles to be overcome through learned abstraction. That insight is not new, but a clean, reproducible demonstration in a well-controlled simulation carries different weight than theory. The question now is whether the robotics field will treat the result as a reason to revisit reservoir computing, or as a curiosity that does not survive contact with the real world.

What the paper does not claim, and what would be easy to overread, is that neuromorphic hardware is about to replace conventional processors in robotics. The Intel Loihi remains a research chip. The control tasks demonstrated are limited. The leap from a simulated arm to a deployable field robot capable of navigating unstructured environments is large and populated with unsolved problems. The 75-times figure is real. The path from simulation to fielded system is not.

The finding is best understood as a directional signal, not a finished verdict: for the specific challenge of soft robot motor control, the AI approach that the field set aside may have been the right one all along. Whether anyone builds on it, and how quickly, depends on whether the robotics research community decides the result is worth taking seriously.