diff --git a/py_workspace/aruco/ground_plane.py b/py_workspace/aruco/ground_plane.py
index ef35040..32a274e 100644
--- a/py_workspace/aruco/ground_plane.py
+++ b/py_workspace/aruco/ground_plane.py
@@ -1,9 +1,9 @@
 import numpy as np
-from typing import Optional, Tuple, List
+from typing import Optional, Tuple, List, Dict, Any
 from jaxtyping import Float
 from typing import TYPE_CHECKING
 import open3d as o3d
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 
 if TYPE_CHECKING:
     Vec3 = Float[np.ndarray, "3"]
@@ -29,6 +29,36 @@ class FloorCorrection:
     reason: str = ""
 
 
+@dataclass
+class GroundPlaneConfig:
+    enabled: bool = True
+    target_y: float = 0.0
+    stride: int = 8
+    depth_min: float = 0.2
+    depth_max: float = 5.0
+    ransac_dist_thresh: float = 0.02
+    ransac_n: int = 3
+    ransac_iters: int = 1000
+    max_rotation_deg: float = 5.0
+    max_translation_m: float = 0.1
+    min_inliers: int = 500
+    min_valid_cameras: int = 2
+
+
+@dataclass
+class GroundPlaneMetrics:
+    success: bool = False
+    correction_applied: bool = False
+    num_cameras_total: int = 0
+    num_cameras_valid: int = 0
+    correction_transform: Mat44 = field(default_factory=lambda: np.eye(4))
+    rotation_deg: float = 0.0
+    translation_m: float = 0.0
+    camera_planes: Dict[str, FloorPlane] = field(default_factory=dict)
+    consensus_plane: Optional[FloorPlane] = None
+    message: str = ""
+
+
 def unproject_depth_to_points(
     depth_map: np.ndarray,
     K: np.ndarray,
@@ -245,3 +275,129 @@ def compute_floor_correction(
     T[:3, 3] = target_normal * t_y
 
     return FloorCorrection(transform=T.astype(np.float64), valid=True)
+
+
+def refine_ground_from_depth(
+    camera_data: Dict[str, Dict[str, Any]],
+    extrinsics: Dict[str, Mat44],
+    config: GroundPlaneConfig = GroundPlaneConfig(),
+) -> Tuple[Dict[str, Mat44], GroundPlaneMetrics]:
+    """
+    Orchestrate ground plane refinement across multiple cameras.
+
+    Args:
+        camera_data: Dict mapping serial -> {'depth': np.ndarray, 'K': np.ndarray}
+        extrinsics: Dict mapping serial -> world_from_cam matrix (4x4)
+        config: Configuration parameters
+
+    Returns:
+        Tuple of (new_extrinsics, metrics)
+    """
+    metrics = GroundPlaneMetrics()
+    metrics.num_cameras_total = len(camera_data)
+
+    if not config.enabled:
+        metrics.message = "Ground plane refinement disabled in config"
+        return extrinsics, metrics
+
+    valid_planes: List[FloorPlane] = []
+    valid_serials: List[str] = []
+
+    # 1. Detect planes in each camera
+    for serial, data in camera_data.items():
+        if serial not in extrinsics:
+            continue
+
+        depth_map = data.get("depth")
+        K = data.get("K")
+
+        if depth_map is None or K is None:
+            continue
+
+        # Unproject to camera frame
+        points_cam = unproject_depth_to_points(
+            depth_map,
+            K,
+            stride=config.stride,
+            depth_min=config.depth_min,
+            depth_max=config.depth_max,
+        )
+
+        if len(points_cam) < config.min_inliers:
+            continue
+
+        # Transform to world frame
+        T_world_cam = extrinsics[serial]
+        # points_cam is (N, 3)
+        # Apply rotation and translation
+        R = T_world_cam[:3, :3]
+        t = T_world_cam[:3, 3]
+        points_world = (points_cam @ R.T) + t
+
+        # Detect plane
+        plane = detect_floor_plane(
+            points_world,
+            distance_threshold=config.ransac_dist_thresh,
+            ransac_n=config.ransac_n,
+            num_iterations=config.ransac_iters,
+        )
+
+        if plane is not None and plane.num_inliers >= config.min_inliers:
+            metrics.camera_planes[serial] = plane
+            valid_planes.append(plane)
+            valid_serials.append(serial)
+
+    metrics.num_cameras_valid = len(valid_planes)
+
+    # 2. Check minimum requirements
+    if len(valid_planes) < config.min_valid_cameras:
+        metrics.message = f"Found {len(valid_planes)} valid planes, required {config.min_valid_cameras}"
+        return extrinsics, metrics
+
+    # 3. Compute consensus
+    try:
+        consensus = compute_consensus_plane(valid_planes)
+        metrics.consensus_plane = consensus
+    except ValueError as e:
+        metrics.message = f"Consensus computation failed: {e}"
+        return extrinsics, metrics
+
+    # 4. Compute correction
+    correction = compute_floor_correction(
+        consensus,
+        target_floor_y=config.target_y,
+        max_rotation_deg=config.max_rotation_deg,
+        max_translation_m=config.max_translation_m,
+    )
+
+    metrics.correction_transform = correction.transform
+
+    if not correction.valid:
+        metrics.message = f"Correction invalid: {correction.reason}"
+        return extrinsics, metrics
+
+    # 5. Apply correction
+    # T_corr is the transform that moves the world frame.
+    # New world points P' = T_corr * P
+    # We want new extrinsics T'_world_cam such that P' = T'_world_cam * P_cam
+    # T'_world_cam * P_cam = T_corr * (T_world_cam * P_cam)
+    # So T'_world_cam = T_corr * T_world_cam
+
+    new_extrinsics = {}
+    T_corr = correction.transform
+
+    for serial, T_old in extrinsics.items():
+        new_extrinsics[serial] = T_corr @ T_old
+
+    # Calculate metrics
+    # Rotation angle of T_corr
+    trace = np.trace(T_corr[:3, :3])
+    cos_angle = np.clip((trace - 1) / 2, -1.0, 1.0)
+    metrics.rotation_deg = float(np.rad2deg(np.arccos(cos_angle)))
+    metrics.translation_m = float(np.linalg.norm(T_corr[:3, 3]))
+
+    metrics.success = True
+    metrics.correction_applied = True
+    metrics.message = "Success"
+
+    return new_extrinsics, metrics
diff --git a/py_workspace/tests/test_ground_plane.py b/py_workspace/tests/test_ground_plane.py
index 9d21beb..0f30a49 100644
--- a/py_workspace/tests/test_ground_plane.py
+++ b/py_workspace/tests/test_ground_plane.py
@@ -5,8 +5,11 @@ from aruco.ground_plane import (
     detect_floor_plane,
     compute_consensus_plane,
     compute_floor_correction,
+    refine_ground_from_depth,
     FloorPlane,
     FloorCorrection,
+    GroundPlaneConfig,
+    GroundPlaneMetrics,
 )
 
 
@@ -315,3 +318,160 @@ def test_compute_floor_correction_bounds():
 
     assert not result.valid
     assert "exceeds limit" in result.reason
+
+
+def test_refine_ground_from_depth_disabled():
+    config = GroundPlaneConfig(enabled=False)
+    extrinsics = {"cam1": np.eye(4)}
+    camera_data = {"cam1": {"depth": np.zeros((10, 10)), "K": np.eye(3)}}
+
+    new_extrinsics, metrics = refine_ground_from_depth(camera_data, extrinsics, config)
+
+    assert not metrics.success
+    assert "disabled" in metrics.message
+    assert new_extrinsics == extrinsics
+
+
+def test_refine_ground_from_depth_insufficient_cameras():
+    # Only 1 camera, need 2
+    config = GroundPlaneConfig(min_valid_cameras=2, min_inliers=10)
+
+    # Create fake depth map that produces a plane
+    # Plane at y=-1.0
+    width, height = 20, 20
+    K = np.eye(3)
+    K[0, 2] = 10
+    K[1, 2] = 10
+    K[0, 0] = 20
+    K[1, 1] = 20
+
+    # Generate points on plane y=-1.0
+    # In camera frame (assuming cam at origin looking -Z), floor is at Y=-1.0
+    # But wait, standard camera frame is Y-down.
+    # Let's assume world frame is Y-up.
+    # If cam is at origin, and looking down -Z (OpenGL) or +Z (OpenCV).
+    # Let's use identity extrinsics -> cam frame = world frame.
+    # World frame Y-up.
+    # So we want points with y=-1.0.
+    # But unproject_depth gives points in camera frame.
+    # If we want world y=-1.0, and T=I, then cam y=-1.0.
+    # But unproject uses OpenCV convention: Y-down.
+    # So y=-1.0 means 1m UP in camera frame.
+    # Let's just make points that form A plane, doesn't matter which one,
+    # as long as it's detected.
+
+    # Let's make a flat plane at Z=2.0 (fronto-parallel)
+    depth_map = np.full((height, width), 2.0, dtype=np.float32)
+
+    # Need to ensure we have enough points for RANSAC
+    # 20x20 = 400 points.
+    # Stride default is 8. 20/8 = 2. 2x2 = 4 points.
+    # RANSAC n=3. So 4 points is enough.
+    # But min_inliers=10. 4 < 10.
+    # So we need to reduce stride or increase size.
+    config.stride = 1
+
+    camera_data = {"cam1": {"depth": depth_map, "K": K}}
+    extrinsics = {"cam1": np.eye(4)}
+
+    new_extrinsics, metrics = refine_ground_from_depth(camera_data, extrinsics, config)
+
+    assert not metrics.success
+    assert "Found 1 valid planes" in metrics.message
+    assert metrics.num_cameras_valid == 1
+
+
+def test_refine_ground_from_depth_success():
+    # 2 cameras, both seeing floor at y=-1.0
+    # We want to correct it to y=0.0
+    config = GroundPlaneConfig(
+        min_valid_cameras=2,
+        min_inliers=10,
+        target_y=0.0,
+        max_translation_m=2.0,
+        ransac_dist_thresh=0.05,
+    )
+
+    width, height = 20, 20
+    K = np.eye(3)
+    K[0, 2] = 10
+    K[1, 2] = 10
+    K[0, 0] = 20
+    K[1, 1] = 20
+
+    # Create points on plane y=-1.0 in WORLD frame
+    # Cam 1 at origin. T_world_cam = I.
+    # So points in cam 1 should be at y=-1.0.
+    # OpenCV cam: Y-down. So y=-1.0 is UP.
+    # Let's just use the fact that we transform points to world frame before detection.
+    # So if we make depth map such that unprojected points + extrinsics -> plane y=-1.0.
+
+    # Let's manually mock the detection to avoid complex depth map math
+    # We can't easily mock internal functions without monkeypatching.
+    # Instead, let's construct a depth map that corresponds to a plane.
+    # Simplest: Camera looking down at floor.
+    # Cam at (0, 2, 0) looking at (0, 0, 0).
+    # World floor at y=0.
+    # Cam floor distance = 2.0.
+    # But here we want to simulate a MISALIGNED floor.
+    # Say we think floor is at y=-1.0 (in our current world frame).
+    # So we generate points at y=-1.0.
+
+    # Let's try a simpler approach:
+    # Create depth map for a plane Z=2.0 in camera frame.
+    # Set extrinsics such that this plane becomes Y=-1.0 in world frame.
+    # Plane Z=2.0 in cam frame: (x, y, 2).
+    # We want R * (x, y, 2) + t = (X, -1, Z).
+    # Let R = Rotation(-90 deg around X).
+    # R = [[1, 0, 0], [0, 0, 1], [0, -1, 0]]
+    # R * (x, y, 2) = (x, 2, -y).
+    # We want Y_world = -1.
+    # So 2 + ty = -1 => ty = -3.
+    # So if we put cam at y=-3, rotated -90X.
+    # Then Z=2 plane becomes Y=-1 plane.
+
+    # Rotation -90 deg around X
+    Rx_neg90 = np.array([[1, 0, 0], [0, 0, 1], [0, -1, 0]])
+    t = np.array([0, -3, 0])
+    T_world_cam = np.eye(4)
+    T_world_cam[:3, :3] = Rx_neg90
+    T_world_cam[:3, 3] = t
+
+    # Depth map: constant 2.0
+    depth_map = np.full((height, width), 2.0, dtype=np.float32)
+
+    # Need to ensure we have enough points for RANSAC
+    # 20x20 = 400 points.
+    # Stride default is 8. 20/8 = 2. 2x2 = 4 points.
+    # RANSAC n=3. So 4 points is enough.
+    # But min_inliers=10. 4 < 10.
+    # So we need to reduce stride or increase size.
+    config.stride = 1
+
+    camera_data = {
+        "cam1": {"depth": depth_map, "K": K},
+        "cam2": {"depth": depth_map, "K": K},
+    }
+    extrinsics = {
+        "cam1": T_world_cam,
+        "cam2": T_world_cam,
+    }
+
+    new_extrinsics, metrics = refine_ground_from_depth(camera_data, extrinsics, config)
+
+    assert metrics.success
+    assert metrics.num_cameras_valid == 2
+    assert metrics.correction_applied
+
+    # We started with floor at y=-1.0. Target is y=0.0.
+    # So we expect translation of +1.0 in Y.
+    # T_corr should have ty approx 1.0.
+    T_corr = metrics.correction_transform
+    assert abs(T_corr[1, 3] - 1.0) < 0.1  # Allow some slack for RANSAC noise
+
+    # Check new extrinsics
+    # New T = T_corr @ Old T
+    # Old T origin y = -3.
+    # New T origin y should be -3 + 1 = -2.
+    T_new = new_extrinsics["cam1"]
+    assert abs(T_new[1, 3] - (-2.0)) < 0.1