MIT Engineers Enable Robots to ‘See’ Hidden Objects Using Generative AI and Penetrating Radar

For more than a decade, a team of researchers at the Massachusetts Institute of Technology (MIT) has pursued a deceptively simple goal: give robots the ability to locate and pick up objects that are concealed behind walls, furniture, or other obstacles. The breakthrough comes from combining low‑power, wall‑penetrating radar with cutting‑edge generative artificial intelligence, allowing machines to reconstruct the shape of invisible items with remarkable precision.

The Challenge of Seeing Through Walls

Robots that operate in cluttered environments—think warehouses, hospitals, or disaster sites—must be able to navigate around unseen hazards and retrieve items that are not directly visible. Traditional vision systems rely on cameras and lasers, which are blocked by opaque surfaces. The idea of using radio waves that can pass through walls dates back to early radar research, but practical implementation has been hindered by limited resolution and noisy data.

From Surface‑Penetrating Radar to Shape Reconstruction

MIT’s original approach employed a single, stationary radar unit that emits low‑power radio waves. When these waves strike a hidden object, they scatter and return to the radar as echoes. By measuring the time delay and frequency shift of each echo, the system can infer a rough, two‑dimensional silhouette of the object. However, this method only provides a coarse outline, leaving large gaps that make it difficult for a robot to determine how to grasp the item safely.

To overcome these limitations, the researchers turned to generative AI. They trained a deep neural network on thousands of paired examples—each consisting of a hidden object and the corresponding radar signature. The model learns to predict the missing portions of an object’s shape based on the partial data it receives, effectively filling in the blanks and producing a complete 3‑D model from a handful of echoes.

Combining Radar with Generative AI

Training the network required a massive dataset of simulated and real radar returns. The team used high‑fidelity physics simulations to generate radar signatures for a wide variety of objects, then validated the results with controlled experiments in a lab environment. The AI’s ability to generalize from these examples means it can handle a broad range of shapes, materials, and orientations, even when the radar data is sparse or noisy.

When the system processes a new radar scan, the generative model first creates a coarse 3‑D reconstruction based on the raw echoes. It then ref