← All robotics laboratories

The book’s brass trash-sorting robot
Robotics & Perception · Section 2.4

The Perception Engine

A sorting robot’s laboratory of likelihood and belief.

INFERENCE LAB№ 024

What does the weight suggest?

Move the weight measurement through the five sensor models. The likelihood compares how well each material explains that same reading.

Weight models

p(weight | category), in g⁻¹

Likelihood at this reading

Raw density values; these need not sum to 1.

Normalized likelihood across weight

Dividing by the sum makes a distribution. It is a posterior only under a uniform prior.

Open the engine: sensor models & assumptions

These are the models in S24_sorter_perception.ipynb. Sensor measurements are conditionally independent given the category. Weight is Gaussian; conductivity and detection use the probabilities below. MAP is a single best category under these assumptions, not a guarantee of the true material.

All inference runs in your browser. Very small likelihoods are combined in log space to preserve numerical accuracy. A zero prior stays zero. Tied maxima are reported together.