Add BIT* planner optimising a pluggable cost function

[H-tr]

What

Adds BIT* (Batch Informed Trees) as a separate planner, following Gammell et al. 2015 and the classic two-queue ompl::geometric::BITstar (OMPL 1.4.2 — OMPL's current BITstar is really Advanced BIT*, which drops the vertex queue).

Unlike RRTConnectPlanner, it returns the cheapest path it found rather than the first feasible one.

The cost interface

CostFunction::state_cost in the new cost_function.hpp is the single stub to fill in. It takes one Config, and reads context() — the vector<Config> of (x, y, theta) dependencies fixed at planner setup:

std::vector<Config> cost_context = {{0.4, 0.1, 0.0}, {0.9, -0.2, 1.57}};
BITStarPlanner planner(&checker, bounds_min, bounds_max, cost_context);
auto path = planner.plan(start, goal, BITStarPlanner::PlanParams{});
const double cost = planner.solution_cost();

It returns 0.0 today, so BIT* currently minimises weighted-SE(2) path length — runnable and testable before the real cost lands.

Why there's a length term

A path is charged the line integral of the density 1 + state_cost. The 1 + is load-bearing:

• it makes heuristic_cost (plain distance) an admissible lower bound, which is what BIT*'s pruning and best-first ordering rest on;
• it keeps every edge cost positive, so rewiring cannot create a cycle;
• it keeps the optimum well defined where state_cost is 0 — including today, with the stub.

A cost that breaks the non-negative contract is clamped in edge_cost rather than left to corrupt the search. This is not theoretical: a state_cost < -1 makes edge costs negative, lets a vertex be reparented under its own descendant, and spins update_descendants() forever. Reproduced, then fixed.

Verification

• 23/23 Catch2 tests pass; 0 failures in 200 repeat runs of [bit_star].
• With state_cost temporarily patched to a Gaussian penalty, the path detours around the costly region (cost 3.0 vs 14.5 straight through) — the cost plumbing genuinely steers the plan.
• With the zero stub, edge_cost == distance bit-for-bit, and solution_cost() equals the summed edge cost of the returned path exactly.
• 40 cluttered-scene trials: zero segment-level collisions, exact endpoints.
• θ-wrapping optimal; start == goal and unreachable-goal cases terminate cleanly.

Also in here

• Bug fix: pyproject.toml used cmake.minimum-version, removed in scikit-build-core 0.8 — pip install -e . was failing outright. (Separate commit.)
• Python surface is additive: Planner.plan_bit_star(), bit_star_solution_cost(), and a cost_context kwarg. Existing plan() is unchanged.
• Renamed run_planner.py → run_rrtc_planner.py (pure rename) and added run_bit_star.py, which sweeps the batch budget so you can watch the cost fall.

🤖 Generated with Claude Code