The noise layer a config asks for with noise_params.

Mirrors tbp/monty/frameworks/models/sensor_modules.py::DefaultMessageNoise. The parameters say how much to move each thing:

noise_params:
  features:
    principal_curvatures_log: [0.1, 0.1]
    pose_vectors: 2.0        # degrees
  location: 0.001            # metres

A real-valued feature gets Gaussian noise of that standard deviation added to it; a boolean flips with that probability; a hue is clipped back into [0, 1] after the noise because there is no colour outside it. A pose_vectors is not nudged component by component — that would stop the three directions being a frame — but turned by a small sampled rotation.

Features are only noised when the percept carries features at all: a step that is a location and nothing else is not worth learning from, and Monty does not pretend otherwise.

The draws are NumPy's and come from the episode's one stream, because in Monty they do: ctx.rng is a single RandomState that every module shares, so a noised percept moves the numbers the motor system draws next. Drawing them from a stream of the sensor's own would leave the episode's stream where it was, and every later draw of the step would then belong to a different experiment.