diff --git a/tests/e2e/test_euroc_vo_only.py b/tests/e2e/test_euroc_vo_only.py
new file mode 100644
index 0000000..957fb33
--- /dev/null
+++ b/tests/e2e/test_euroc_vo_only.py
@@ -0,0 +1,135 @@
+"""VO-only diagnostic for EuRoC MH_01.
+
+Bypasses the full pipeline (ESKF, satellite matching, recovery) and drives
+ORBVisualOdometry directly on raw EuRoC frames. Reports the per-frame tracking
+success rate so we can distinguish:
+
+  - ORB can track EuRoC frames  → vo_success=0 in the full pipeline means the
+    ESKF/satellite layer is the problem, not VO itself.
+  - ORB cannot track EuRoC frames → VO backend is the root cause; need to swap
+    to a real SuperPoint+LightGlue model or cuVSLAM.
+
+Skipped if EuRoC MH_01 is not installed.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import cv2
+import numpy as np
+import pytest
+
+from gps_denied.core.vo import ORBVisualOdometry
+from gps_denied.schemas import CameraParameters
+from gps_denied.testing.datasets.euroc import EuRoCAdapter
+
+# Match CI-tier frame cap used in test_euroc.py
+EUROC_MH01_MAX_FRAMES = 100
+
+# EuRoC cam0 intrinsics from the dataset sensor.yaml (focal length ≈ 458.654 px)
+# Expressed as physical params so CameraParameters can build K.
+# sensor_width chosen so focal_length * (res_w / sensor_width) ≈ 458 px.
+EUROC_CAM_PARAMS = CameraParameters(
+    focal_length=4.586,      # mm (synthetic — tuned so f_px ≈ 458 px)
+    sensor_width=4.586,      # mm → f_px = 4.586 * (752/4.586) = 752... no
+    sensor_height=4.586,
+    resolution_width=752,
+    resolution_height=480,
+    # principal_point omitted → uses image centre (376, 240)
+)
+
+# Actually compute correct params: f_px = focal_length * (res_w / sensor_w)
+# We want f_px = 458.654. Set focal_length=458.654, sensor_width=752 so that
+# f_px = 458.654 * (752/752) = 458.654.
+EUROC_CAM_PARAMS = CameraParameters(
+    focal_length=458.654,
+    sensor_width=752.0,
+    sensor_height=480.0,
+    resolution_width=752,
+    resolution_height=480,
+)
+
+# Minimum acceptable VO tracking rate to call the integration healthy.
+MIN_TRACKING_RATE = 0.70  # 70 % of consecutive frame pairs
+
+
+@pytest.mark.e2e
+@pytest.mark.needs_dataset
+@pytest.mark.asyncio
+async def test_euroc_mh01_orb_tracking_rate(euroc_mh01_root: Path):
+    """ORB VO should track ≥70 % of consecutive EuRoC frame pairs.
+
+    If this xfails it means ORB itself cannot engage on EuRoC imagery — the
+    root cause of vo_success=0 in the full pipeline is the VO backend, not the
+    ESKF or satellite layer.
+    """
+    adapter = EuRoCAdapter(euroc_mh01_root)
+    frames = list(adapter.iter_frames())[:EUROC_MH01_MAX_FRAMES]
+
+    vo = ORBVisualOdometry()
+    success_count = 0
+    total_pairs = 0
+    prev_image: np.ndarray | None = None
+
+    for frame in frames:
+        img = cv2.imread(frame.image_path, cv2.IMREAD_GRAYSCALE)
+        if img is None:
+            continue
+        if prev_image is not None:
+            pose = vo.compute_relative_pose(prev_image, img, EUROC_CAM_PARAMS)
+            total_pairs += 1
+            if pose is not None and pose.tracking_good:
+                success_count += 1
+        prev_image = img
+
+    tracking_rate = success_count / max(1, total_pairs)
+
+    # Always report numbers — visible in pytest -v output even on xfail.
+    summary = (
+        f"ORB tracking: {success_count}/{total_pairs} pairs "
+        f"({tracking_rate*100:.1f}%)"
+    )
+
+    if tracking_rate < MIN_TRACKING_RATE:
+        pytest.xfail(
+            f"{summary} — below {MIN_TRACKING_RATE*100:.0f}% threshold. "
+            "ORB struggles on EuRoC indoor imagery; upgrading to real "
+            "SuperPoint+LightGlue or cuVSLAM is required to make the full "
+            "pipeline viable on this dataset."
+        )
+
+    assert tracking_rate >= MIN_TRACKING_RATE, summary
+
+
+@pytest.mark.e2e
+@pytest.mark.needs_dataset
+@pytest.mark.asyncio
+async def test_euroc_mh01_orb_reports_nonzero_inliers(euroc_mh01_root: Path):
+    """Sanity: ORB must find at least some inlier matches on at least one pair.
+
+    If *every* pair has zero inliers the images are likely being loaded
+    incorrectly (wrong path, wrong colour mode, etc.).
+    """
+    adapter = EuRoCAdapter(euroc_mh01_root)
+    frames = list(adapter.iter_frames())[:20]  # small — just a sanity check
+
+    vo = ORBVisualOdometry()
+    any_inliers = False
+    prev_image: np.ndarray | None = None
+
+    for frame in frames:
+        img = cv2.imread(frame.image_path, cv2.IMREAD_GRAYSCALE)
+        if img is None:
+            continue
+        if prev_image is not None:
+            pose = vo.compute_relative_pose(prev_image, img, EUROC_CAM_PARAMS)
+            if pose is not None and pose.inlier_count > 0:
+                any_inliers = True
+                break
+        prev_image = img
+
+    assert any_inliers, (
+        "ORB found zero inliers across the first 20 EuRoC frame pairs. "
+        "Check that image paths are correct and images load as valid grayscale arrays."
+    )