#!/usr/bin/env python3 """Track A, stage 1: score every camera view of every scene with Cosmos-Reason2 (CPU ok). Writes one JSON line per (scene, camera) to results/trackA/vlm_scores.jsonl and resumes from it, so the expensive VLM pass runs once; all fusion/attack experiments reuse it. """ import argparse import json import sys import time from pathlib import Path sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "src")) from robotics.sim.tabletop import CAMERAS, make_suite # noqa: E402 from robotics.vlm.cosmos_reason import CosmosVerifier # noqa: E402 def main(): ap = argparse.ArgumentParser() ap.add_argument("--n", type=int, default=50) ap.add_argument("--seeds", type=int, nargs="+", default=[100, 200]) ap.add_argument("--out", default="results/trackA/vlm_scores.jsonl") a = ap.parse_args() out = Path(a.out) out.parent.mkdir(parents=True, exist_ok=True) done = set() if out.exists(): for line in out.read_text().splitlines(): r = json.loads(line) done.add((r["seed"], r["scene"], r["camera"])) v = CosmosVerifier() with out.open("a") as f: for seed in a.seeds: suite = make_suite(a.n, seed=seed, cameras=CAMERAS) for i, s in enumerate(suite): for cam in CAMERAS: if (seed, i, cam) in done: continue t = time.perf_counter() p = v.score(s.views[cam], s.task) f.write(json.dumps({"seed": seed, "scene": i, "camera": cam, "task": s.task, "kind": s.kind, "gt": s.success, "p": round(p, 5), "latency_s": round(time.perf_counter() - t, 3)}) + "\n") f.flush() print(f"seed {seed} scene {i + 1}/{len(suite)}", flush=True) if __name__ == "__main__": main()