The random walk InformedPolicy does, on its own.

Mirrors tbp/monty/frameworks/models/motor_policies.py::InformedPolicyRandomWalk. InformedPolicy is a walk with a recovery rule: sample while the sensor is on the object, and step back while it is not. This policy is that rule without the rest of InformedPolicy, which is what the distant agent's motor-system configs ask for as their selector's default (distant_2, distant_5): the selector routes goals to JumpToGoal and LookAtGoal, and between goals the body wanders.

There is no momentum here, whatever the word "random walk" suggests: every step is an independent draw from the sampler. Monty's own docstring says as much — this reproduces what InformedPolicy did rather than proposing a better walk — and the alpha momentum belongs to SurfacePolicy's tangential direction, not to this policy.

The remembered action is inverted every time it is used, so a sensor that keeps leaving the object walks back and forth across the edge instead of away.