MIT and WHOI System Lets ROVs See Through Underwater Sediment Clouds

Researchers at the Woods Hole Oceanographic Institution have built a system that lets remotely operated underwater vehicles navigate through the sediment clouds they stir up when touching down on the seafloor. Developed by Amy Phung (SM '23, PhD '26) and her advisor Richard Camilli (SM '00, PhD '03), the technique combines sonar mapping with a real-time depth-estimation algorithm to guide a vehicle safely through zero-visibility conditions, then bring it close enough for onboard cameras to take over.

What happened

Detail	Fact
Lead researcher	Amy Phung, SM ’23, PhD ’26
Advisor	Richard Camilli, SM ’00, PhD ’03
Institution	Woods Hole Oceanographic Institution (WHOI)
Core sensor	Sonar for initial mapping
Algorithm source	Image-matching depth estimator developed by researchers in France
Key capability	Real-time 3D mapping in turbid (murky) water

Remotely operated vehicles (ROVs) often land on the seafloor or dig through sand beds during underwater missions. That movement kicks up sediment clouds that cameras cannot see through, and the standard fix is simply to wait. Phung and Camilli designed a way around that delay.

The system works in two stages. First, the ROV uses sonar to build a quick map of its surroundings. Sonar works equally well in clear or murky water, though it does not deliver the fine detail that cameras can. The map is enough to steer the vehicle safely toward a specific object without colliding with anything nearby.

Once the vehicle is close enough, the cameras can resolve the object clearly because the sediment cloud is thinner at short range. The researchers accelerated the whole process by pairing the sonar data with a pixel-level depth estimation algorithm built by a team in France. That algorithm infers the relative distance of every pixel in a 2D frame, effectively adding a depth dimension to what is otherwise a flat image, fast enough to run in real time.

Camilli described the challenge in plain terms: “An analogy would be if you were to go into a china shop in the dark, and try to pick your way around to find a specific coffee mug without knocking things over. This would allow you to do that.”

Why it matters

Phung and Camilli say the system could serve several practical areas: scientific seafloor exploration, underwater construction and maintenance, and the detection and handling of unexploded undersea mines. Each of those involves working in conditions where sediment disturbance is common and waiting is either costly or dangerous.

For industries that deploy ROVs regularly, reducing downtime caused by visibility loss has direct operational value. A vessel and its crew waiting on the surface while sediment settles is time and money gone. A system that keeps the vehicle productive during that window matters more than it might first appear.

The broader pattern here is also worth noting: combining a low-resolution but robust sensor (sonar) with a higher-resolution but fragile one (cameras), bridged by a fast inference algorithm, is an approach that keeps appearing across robotics and AI integration projects in harsh physical environments.

Our take

This is solid applied research, not a product announcement. The gap between a working lab prototype and a commercially deployed ROV system is real, and the paper does not claim otherwise. What is genuinely interesting is the architecture: sonar for coarse navigation, depth estimation for spatial awareness, cameras for fine inspection. Each sensor does what it is actually good at.

The French depth-estimation algorithm doing the heavy lifting in real time is a detail worth watching. Monocular depth estimation (inferring depth from a single 2D image rather than a stereo pair) has improved significantly in recent years, and this is one of the more concrete demonstrations of it working outside a controlled lab setting. If you follow on-device AI benchmarking, the real-time constraint here is a meaningful bar to clear.

For businesses, the direct application is narrow unless you run subsea operations. But the underlying principle, using fast AI-assisted inference to extend sensor usefulness in degraded conditions, is showing up in drone navigation, autonomous vehicles, and industrial inspection. It is worth understanding now.

Source: MIT Technology Review