diff --git a/data_only_viz/scripts/bench_icp_fusion.py b/data_only_viz/scripts/bench_icp_fusion.py
new file mode 100644
index 0000000..b89c3cd
--- /dev/null
+++ b/data_only_viz/scripts/bench_icp_fusion.py
@@ -0,0 +1,77 @@
+"""Latency / convergence bench for the ICP fusion worker.
+
+Usage:
+
+    cd data_only_viz
+    uv run --extra lidar python -m data_only_viz.scripts.bench_icp_fusion \
+        --n-frames 200 --n-people 2 --seed 0
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import time
+
+import numpy as np
+
+from data_only_viz.icp_fusion import FusionWorker, IcpConfig
+from data_only_viz.lidar_calib import Extrinsic
+from data_only_viz.state import SMPLXPerson, State
+
+
+def _synth_person(seed: int, offset_x: float) -> SMPLXPerson:
+    rng = np.random.RandomState(seed)
+    verts = np.zeros((10475, 3), dtype=np.float32)
+    pts = rng.randn(2000, 3).astype(np.float32) * 0.1
+    verts[: pts.shape[0]] = pts + np.array([offset_x, 0, 1.5], dtype=np.float32)
+    verts[5559] = pts.mean(axis=0) + np.array([offset_x, 0, 1.5], dtype=np.float32)
+    return SMPLXPerson(pid=seed, vertices_3d=verts)
+
+
+def main(argv: list[str] | None = None) -> int:
+    p = argparse.ArgumentParser()
+    p.add_argument("--n-frames", type=int, default=200)
+    p.add_argument("--n-people", type=int, default=2)
+    p.add_argument("--seed", type=int, default=0)
+    args = p.parse_args(argv)
+
+    rng = np.random.RandomState(args.seed)
+    persons = [_synth_person(i, offset_x=-0.6 + 1.2 * i) for i in range(args.n_people)]
+    state = State()
+    state.persons_smplx = persons
+
+    worker = FusionWorker(extrinsic=Extrinsic.identity(), config=IcpConfig())
+
+    latencies_ms: list[float] = []
+    accepted = 0
+    pelvis_delta_m: list[float] = []
+    for _ in range(args.n_frames):
+        all_pts = np.concatenate([
+            pers.vertices_3d[: 2000] + np.array([0, 0.05, 0], dtype=np.float32) +
+            0.02 * rng.randn(2000, 3).astype(np.float32)
+            for pers in persons
+        ])
+        state.lidar_points = all_pts
+        before = np.stack([p.vertices_3d[5559].copy() for p in state.persons_smplx])
+        t0 = time.perf_counter()
+        meta = worker.run_once(state)
+        latencies_ms.append((time.perf_counter() - t0) * 1000.0)
+        accepted += len(meta.applied)
+        after = np.stack([p.vertices_3d[5559] for p in state.persons_smplx])
+        pelvis_delta_m.extend(np.linalg.norm(after - before, axis=1).tolist())
+
+    report = {
+        "n_frames": args.n_frames,
+        "n_people": args.n_people,
+        "latency_ms_p50": float(np.percentile(latencies_ms, 50)),
+        "latency_ms_p95": float(np.percentile(latencies_ms, 95)),
+        "acceptance_rate": accepted / (args.n_frames * args.n_people),
+        "pelvis_delta_m_mean": float(np.mean(pelvis_delta_m)),
+        "pelvis_delta_m_max": float(np.max(pelvis_delta_m)),
+    }
+    print(json.dumps(report, indent=2))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())