Planner and sim speed-ups, trajsample overhaul, motion metrics

[isaac0804]

Speed-ups for the planner and simulator, a trajectory-sampling overhaul, motion and collision metrics, and a per-strategy preferred control scheme.

What's in it

Simulator and startup

• rSim (vendor/rSim): a raw step buffer (step_fast, get_state_array) and a Release -O3 -march=native build. pybind11 is bumped to 2.13.6, with its hash pinned. robosim_wrapper falls back to the old calls when the extension doesn't have them, so CI's rc-robosim is unaffected.
• The robosim subprocess launches .pixi/envs/robosim/bin/python directly instead of through pixi run, cutting startup from about 469 ms to 19 ms. The sim is reused across matches within a tournament.
• Frame generation avoids redundant dict copies.

FastPathPlanner

• Optional C++ extension (fastpathplanning/cpp, built with pixi run build-planner). It's used when present; the numba kernels are used otherwise. Exact-math mode stays bit-identical (fastpathplanner_equivalence_test.py). The C++ fast path matches numba to 1e-9 (test_fastpathplanner_cpp.py), and that module skips when the extension isn't built (it isn't built in CI).

Trajectory sampling

• Fixes over-inflation of the transverse reachability tube in collision_numba.
• Picks the candidate with the shortest arrival time, with 0.15 s hysteresis.
• A symmetric fan of intermediate targets whose radius scales with distance to the target.
• Sumatra-style fallback ranking, and role-aware priority (the ball carrier outranks off-ball teammates).
• The motion planning benchmark grows to 29 scenarios; trajsample passes all of them.

Metrics and control scheme

• motion_stats and MatchStats: robot and teammate overlap ticks, near-miss ticks, worst clearance, and discrete collision incidents, reported in the round-robin summary.
• Strategies can declare a preferred control scheme (AbstractStrategy.control_scheme). round_robin's default (fpp) is unchanged, and no strategy declares one yet.

Caveats

• Trajsample is not ready to be the default. The 2026-10-14 trajsample round-robin (tournament_20261014_105956) had 10 of 66 matches stall and 248 goals. The 2026-10-10 fpp run on the same 66 pairings had no stalls and 471 goals.
• The C++ FPP fast path can change match results. Whether the extension is built affects results within the 1e-9 tolerance, and the match cache doesn't record it. Round-robins meant to be compared should be run with the same build.
• The sim and planner changes alter the round-robin fingerprint, so every cached match reruns. A new fpp baseline will be run after this and Real SSL logs as replays, goal ledger, situation bench, and referee fixes #143 are merged.

Tests

• origin/main + this branch + Real SSL logs as replays, goal ledger, situation bench, and referee fixes #143 merged: lint passes, and pytest utama_core/tests/ --headless --level quick --ignore-glob "**/*grsim*" gives 12680 passed, 87 xfailed, 0 failed.
• Without the extension, utama_core/tests/motion_planning/ gives 1612 passed, 1 skipped (the C++ module), 87 xfailed. With it built, test_fastpathplanner_cpp.py passes 9 of 9.

🤖 Generated with Claude Code