Generative AI improves a wireless vision system that sees through obstructions

Source Domain: news.mit.edu

MIT researchers have developed new techniques for robots to “see” through obstacles to locate and manipulate hidden objects using surface-penetrating wireless signals.
They used generative artificial intelligence (AI) to fill in gaps in reflections and achieved more accurate reconstructions of hidden objects and entire rooms.
The approach uses millimeter wave signals similar to Wi-Fi to penetrate materials, but applies AI to interpret the specular reflections that typically obscure hidden objects.
They trained generative AI models on synthetic datasets created by adapting existing computer vision images to mimic wireless reflections, to overcome the lack of large existing datasets.
The system, called Wave-Former, achieved nearly 20% better accuracy reconstructing hidden objects compared to existing techniques.
The researchers also developed RISE, a system that reconstructs whole rooms by analyzing “ghost signals” from reflections of human motion, and achieved more precise scene reconstructions.
These innovations could enable applications like verifying packed items in warehouses, or allowing robots to understand human locations for improved human-robot interaction.
Future work includes building foundation models for wireless signals to enable a wider range of applications.

Generative AI improves a wireless vision system that sees through obstructions | MIT News