Images, sent as vibration. Two copies at once, recovered from the sound alone.
Radio fades within meters of seawater, and light within tens of meters at best. Sound carries for kilometers. Vibrational turns a picture into a few seconds of sound, sends it twice at once in two frequency bands, and recovers the image on the other side.
Fidelity 34.9 dB PSNR · SSIM 0.926
Color never sent · inferred by AI
Real output of the Vibrational pipeline on a validation photo the network never trained on, played back as transmitted. Each lane traces the strength of one copy over time, from the recording itself; the individual waves are drawn larger than life so they are visible. Clean software channel; not an in-water measurement. Lower your volume: it is a dense tone.
A burst damages a different part of the photo in each copy.
The problem. Real channels are noisy. A passing ship, a knock on a pipe or a burst of static wipes out a moment of the signal, and with it part of the picture.
The fix. The second copy runs backwards. When the burst hits both bands at once, it erases the left of the photo in the forward copy but the right of the photo in the reversed copy. Every part survives in at least one copy, at no cost in airtime.
The result. Keeping, column by column, whichever copy looks less noisy recovers the photo at 34.9 dB, the same as with no burst at all. A steady tone sitting in one band would likewise spoil only one copy.
A real 0.77-second burst of random noise added to the recording, then decoded by the Vibrational receiver. Scores are PSNR against the photo that was sent. The per-column choice uses only the received copies, never the original; the main pipeline currently averages the two copies, which suits steady noise but not a burst like this one.
The Voyager recently achieved a massive milestone: traveling a full light day away from our home planet, with communication still intact.
However, similar milestones just beneath our feet, in oceans and subterranean environs, have barely been scratched. This is in large part due to the challenge of transmitting signals through media thicker than air.
While space has often been hailed as the “final frontier,” the frontiers here at home, on our planet, have often taken a back seat. However, as technology and robotics have advanced in autonomy and capability, these arenas are being revisited. From defense to environmental research, the ability to communicate underwater and even deep underground is becoming ever more important.
The purpose of this project, Vibrational, is to use lower frequencies to battle this challenge. Fans of baseball may be familiar with the slowness with which sound travels through air. At 150 meters from home plate, the sound of the bat connecting with the ball arrives half a second after your eyes perceive it. Imagine if cell signal took that long to reach your phone from the nearest tower.
Underwater and underground, this paradigm is reversed. The thicker media transmit sound with higher velocity and fidelity, while visible light and other electromagnetic waves are obscured in the opposite way.
The goal is to leverage this to create acoustic and even seismic data transmission systems that take advantage of this gift of physics. While traditional acoustic modems are useful, they often lack the ability to transmit data at speed in the presence of acoustic interference. Next-generation approaches allow this paradigm to evolve.
As a human, if you are having a conversation on a subway platform about what is happening on the other side of the tracks and a noisy train goes by, the other person’s voice and the image on the other side of the tracks do not just become blackness and deafness.
To be sure, it is obstructed, but not incomprehensible. Similarly, if you were by some “magic” able to slow down time, you could most likely make out exactly what is being said and what is happening.
In practice there is rarely just one train. One passes left to right, another right to left, and both of you are talking at once. Yet what one train hides, the other usually leaves open a moment later. Vibrational builds this into the signal itself: every image is sent twice at the same moment, once forward and once backwards, so a burst of noise seldom hides the same part of the picture in both copies. And unlike a person on the platform, a receiver really can slow time down. It holds the whole recording and can study every instant of it before deciding what was said.
That, in principle, is the goal. Not to achieve perfection when perfection is impossible. It is to transmit data such as images, videos, conversation and text in a way that allows intelligent machines to “intuitively” reconstruct the degraded signals.
In doing so, we may yet achieve perfect transmission in hostile environments, ironically by relaxing the relentless pursuit of it, much the way human perception does.
One picture, two voices, one machine that fills in the gaps.
Strip it to light and shadow
The photo becomes grayscale. Structure travels; color stays behind, to be inferred later.
Speak it twice, at once
The image is sent as sound in two separate frequency bands at the same moment, one copy played backwards. Redundancy without extra airtime.
Listen, align, combine
The receiver decodes each band on its own, turns the reversed copy back around, and averages the two into one image.
Let a network see color
A neural network trained on 12,928 photos infers color from the received grayscale, the way your brain fills in a dim room. The colors are its best guess, not a recovery.
Measured so far.
Airtime for one 256×256 image, including synchronization. The second copy adds no time.
Mean PSNR of the recovered image against the grayscale that was sent.
Mean structural similarity (SSIM), where 1.0 is identical.
Photographs transmitted and recovered to produce these averages.
Each row shows the grayscale that was transmitted, the image recovered from the received sound, and the network's colorized version. The scores compare the recovered image with the transmitted grayscale. No color is sent: the colorized images are an early model's inference and can differ from the original photo's colors. These are selected examples from a 1,624-photo validation set, chosen to show the system at its best; across the whole set the recovered image averages 29.0 dB. All results use a clean, noise-free software channel, and no in-water or seismic results are claimed. Images are unedited pipeline outputs. Photographs: USIS16K dataset.
From a clean channel to a hostile one.
Dual-band transmission
Encoder and receiver run end to end on real photographs, with tests.
Reconstructing through noise
Neural networks that restore detail from deliberately degraded transmissions, then a sealed final test.
Color from structure
A diffusion-based network that infers full color from the received grayscale, since color is never sent.
Water and rock
In-water acoustic trials, then seismic links, and from still images to video, voice and text.
Andrew Cavallaro
Machine learning and robotics engineer working on perception and communication in degraded, bandwidth-limited environments, especially underwater. Based between New York and Rome; MS in AI & Robotics at Sapienza University of Rome.
Open to roles in underwater robotics, defense and ML research.