Version 1 record

This page describes software that has been retired. It is accurate about Version 1 and is not a description of what runs now — see the current site and the archive.

C-15 Planner

Status: built

Code: services/planner/ — value.py, collapse.py for the simulated collapse of uncertainty along a route, select.py for the prize-collecting selection, projection.py for the forward growth, and commitment.py for the receding horizon
Delivered by: specs/011-adaptive-planner
Covered by: services/planner/tests/, including test_optimality_gap.py and test_selection_determinism.py, plus tests/integration/test_planner_replans_on_new_field.py and tests/integration/test_plan_carries_no_instruction.py, which asserts the boundary described below rather than trusting it

Responsibility: adaptive sampling recommendations. Owns the failure mode of: crossing into tactical advice.

What it does

It reads the uncertainty field and works out where sampling would reduce it most, then emits a route. The interesting part is that it simulates the collapse of uncertainty along each candidate route as that route is traversed, so the value of a distant objective decays as nearer sampling resolves the same thing. Without that, a planner recommends every high-uncertainty cell independently and produces a route that visits the same information twice.

It replans on a receding horizon as measurements arrive, and it projects uncertainty growth forward so that it can report when a region will fall below usable confidence — which makes the output schedulable rather than merely reactive.

The route selection problem is treated as orienteering, or prize-collecting: a budget is spent on the cells worth visiting. It is explicitly not a travelling-salesman problem, because nothing requires every cell to be visited and treating it that way produces long routes that acquire nothing.

The boundary it defends

The planner emits recommendations. It does not command, does not task, and does not advise a human directly. Rendering and advice happen downstream.

This is a line worth stating precisely, because it is easy to state loosely. Computing where sampling would most reduce uncertainty is decision logic, even when nothing ever draws it. The boundary being defended is not "does the system decide" — it does — but who recommends: the system produces a recommendation, and a human elsewhere decides what to do with it.

Requirements: FR-32 to FR-36. Feature: 011. See also the derivation of informative path planning.