RIAT LABS
28 October 2026 · 2 min readprojectedge AIsensor fusion

Sanatorium Ingest Photobox: measuring food waste with edge AI

A multi-sensor capture unit for food-waste analysis, built in the aws Food Explore project: stereo 3D, weight and time-of-flight sensing with AI that runs locally.

Abstract plate with a detection box

Care and hospital kitchens throw away a surprising share of what they cook, and nobody can say exactly how much, or of what. The Sanatorium Ingest Photobox is our answer to the measurement problem: a self-contained unit that looks at a plate or tray, weighs it, and lets an AI model work out what was left. The project is funded by aws Food Explore (project P2504754).

Why sensor fusion

A photo alone tells you what is on the plate, but not how much. A scale alone tells you how much, but not what. Combining the two, and adding depth, turns guesses into numbers.

Simplicity is prerequisite for reliability.

— Edsger W. Dijkstra, EWD498

How it fits together

The Photobox is an edge-AI capture unit. All sensing and inference happen inside the box.

Photobox sensors feed an edge unit that runs local AI and produces a waste report
Figure 1: simplified architecture of a multi-sensor capture unit (illustration)

SensingTypical componentContributes
3D / stereo captureCamera pairShape and volume of the food
WeightPrecision scale (KERN class)Mass before and after
Depth and distanceLiDAR, time-of-flightCross-check of volume and position
Compute and I/ORaspberry Pi Compute Module 4, ESP32-S3Capture, control and inference

A look at the numbers

The core of the analysis is simple arithmetic on fused measurements. This sketch shows the idea; the real pipeline works on 3D data and per-dish models.

def food_waste_g(tray_before_g: float, tray_after_g: float, tray_empty_g: float) -> float:
    """Waste = what is still on the tray after the meal."""
    return max(0.0, tray_after_g - tray_empty_g)

def waste_share(before_g: float, after_g: float, empty_g: float) -> float:
    served = before_g - empty_g
    return food_waste_g(before_g, after_g, empty_g) / served if served > 0 else 0.0

Names such as waste_share() are styled inline as well.


What we are working on

  • Reliable volume estimation from stereo and depth data.
  • Models that are small enough for the edge and still tell dishes apart.
  • A capture routine that works in a busy kitchen, not only on the bench.

Food Explore is part of our wider work on edge AI and sensor fusion.

← All posts