top of page
PATENT & INVENTION
Patent Application: NIR-Based Navigation Device
Filed independently in 2026 as sole inventor, addressing a gap left by conventional RGB-based robot vision in disaster environments.
NIR-Based Navigation Device with a Knowledge
- Distilled Embedded Neural Network
2026 · Independent invention, patent pending
ABSTRACT
This invention proposes a navigation device for visibility-limited environments — smoke, fog, dust, and low light — where conventional RGB camera vision fails. It combines a near-infrared (NIR) camera with active illumination and a neural network compressed via knowledge distillation to run entirely on a low-power microcontroller, without cloud computing.
BACKGROUND
Conventional robot vision relies on RGB cameras, but visible light scatters or is blocked in smoke, fog, and darkness, leaving robots unable to recognize obstacles or navigate. Large AI vision models can handle these conditions accurately, but their size and compute demands make them impossible to run on the small, low-power processors found in compact disaster-response robots — creating a real gap between what's accurate and what's actually deployable in the field.
CLAIMED INVENTION
The system pairs a 940nm NIR camera and active illuminator (capturing object outlines through smoke and darkness) with a preprocessing stage (contrast enhancement, resizing, INT8 quantization) and an embedded inference engine. That engine runs a neural network trained through knowledge distillation — compressed from a ~410,000-parameter teacher model down to a 5,988-parameter student model, quantized to just 15.7 KB — small enough to run in real time on a microcontroller with only 1.5 MB of memory. Classification results are converted directly into drive commands (forward, left, right, backward), with a built-in safety module that halts the robot whenever prediction confidence drops below a set threshold.
CLAIMED RESULTS
The compressed student model retained approximately 86% of the teacher model's accuracy despite a roughly 100-fold reduction in model size — confirming that knowledge distillation could preserve enough perceptual accuracy for real-world navigation while fitting entirely on embedded, cloud-free hardware.
FIELD OF APPLICATION
A perception system this small and this cheap changes what's possible for disaster-response robotics: it opens the door to low-cost, reusable navigation hardware that can assist firefighters entering smoke-filled buildings or guide small search-and-rescue robots — without needing an internet connection, a powerful onboard computer, or expensive sensors. It's a working example of how model compression can turn a lab-grade capability into something field-deployable.
bottom of page