Vibrations and AI Could Show Robots What Lies Beneath Alien Terrain

In a test bed of sand, a researcher swings an instrumented hammer and strikes a block of concrete poking out of the surface. A laser, aimed from a distance, picks up the tiny tremors that ripple across the block. Those tremors, it turns out, can say how much of the block is hidden below.

That is the core of new work from Virginia Tech, published in IEEE Sensors Journal. Erik Komendera, an assistant professor of mechanical engineering who once worked as an engineer at NASA Langley Research Center, led it with Ph.D. candidate Yiyan Ruan. Their aim is a set of tools that tell a robot how deep a rock is buried, information that could help rovers land safely on Mars, the moon or other largely unexplored worlds.

The problem is practical. On current Mars missions, the elevation and roughness of the terrain and the height of rocks all shape where a spacecraft can be set down. A rock sitting in the way of a landing surface might be something a robotic excavator can chip away. But if what looks like a rock is really the top of a buried boulder, the machine would waste energy and wear down its tools on something it cannot move.

Komendera framed it as an old human problem. For as long as people have built things on Earth, he said, they have had to deal with large, heavy obstacles that can't be ignored or scooped aside without planning. A person can look at a rock and quickly judge whether it is resting or partly buried, and whether it can be moved alone or needs help and tools. The question was how a robot could make the same judgment. "Now we have an answer," he said.

To test the approach, the team buried 18 concrete blocks of different sizes and shapes at different depths in sand. They struck the exposed part of each block with the instrumented hammer and measured the resulting vibrations remotely with a scanning laser Doppler vibrometer. They repeated the measurements at multiple heights on the surface of each block.

The readings were then fed to a neural network, which learned the relationship among three things: the vibration characteristics, the visible geometry of the rock, and its burial depth. The task was not simple. The readings are affected by soil stiffness, the angle at which a rock meets the sand, and other factors. By training the AI to account for them, the team built models that could reliably determine depth.

Ruan explained the difficulty. Estimating burial depth from vibrations can be tricky, she said, because the feedback depends on the rock's geometry and on how the soil holds it. By carefully selecting features from the vibration measurements and the exposed geometry, the team got reliable results from a compact neural network with just two hidden layers.

The models could one day guide a robotic excavator on Mars in handling obstacles correctly. For now, the work stays on Earth. The team plans to keep building datasets and testing the method here, where it could help determine the components of unstable rock formations, measure underground caverns and assist with disaster cleanup, in settings where hazards are buried. The next step is to see how well the method holds up beyond concrete blocks in sand.