Results¶
Simulation only: these rank digital twins, not yet the real cameras
Every catalog lens is a surrogate (a published design scaled to the camera's focal length and f-number; vendors publish no prescriptions), and sensor noise is not simulated. The main tables put each camera on the robot's wrist; the third-person tables below them are secondary (LIBERO's agent-view camera is not on the robot). Grounding partly reflects how the model handles each image's size and aspect. See Findings.
Benchmark commit da98e26 · TwinRobo-Core 075020f, 48bad68 · copied 2026-10-12.
On the robot: each camera replacing the robot's wrist camera¶
Where a robot's camera sits: on the wrist, close above the table and moving with the arm. Field of view, resolution, lens and focus together. The board columns are the camera's own (the board test does not depend on the mount).
| Camera | Status | Image (px) | HFoV (deg) | Board range (m) | Board found (%) | Pose error (% of dist.) | Corner error (mrad) | Detection mAP50 | Segmentation mIoU | Object depth AbsRel, DA-V2 (%) | VLM grounding hit (%) |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Ideal pinhole (1024x1024, 75 deg) | ideal | 1024x1024 | 75 | – | – | – | – | 39.5 | 94.3 | 9.1 | 62.8 |
| Intel RealSense D435 / D435i RGB camera | estimated | 1920x1080 | 69 | 0.15–1.99 | 64 | 0.02 | 0.074 | 33.5 | 93.1 | 8.3 | 59.8 |
| Intel RealSense D455 RGB camera | estimated | 1280x800 | 90 | 0.15–0.91 | 44 | 0.06 | 0.240 | 36.2 | 94.0 | 9.7 | 61.4 |
| Intel RealSense D435 / D435i IR imager (one stereo eye) †* | estimated | 1280x800 | 90 | 0.15–1.01 | 45 | 0.05 | 0.220 | 33.8 | 94.3 | 9.5 | 52.8 |
| Intel RealSense D455 IR imager (one stereo eye) †* | estimated | 1280x800 | 90 | 0.15–1.01 | 45 | 0.05 | 0.220 | 33.8 | 94.3 | 9.5 | 52.7 |
| Logitech C920 HD Pro Webcam | estimated | 1920x1080 | 70 | 0.15–1.78 | 66 | 0.03 | 0.099 | 35.6 | 92.7 | 8.6 | 62.2 |
| Luxonis OAK-D color camera | estimated | 4056x3040 | 69 | 0.15–1.99 | 70 | 0.02 | 0.046 | 38.3 | 93.2 | 9.7 | 61.0 |
| Luxonis OAK-D mono camera (one stereo eye) * | estimated | 1280x800 | 72 | 0.15–1.27 | 56 | 0.02 | 0.086 | 30.3 | 93.9 | 9.2 | 51.5 |
| Raspberry Pi Camera Module 3 | estimated | 4608x2592 | 66 | 0.21–3.12 | 84 | 0.05 | 0.106 | 35.8 | 92.7 | 8.9 | 56.9 |
| Raspberry Pi Camera Module 3 Wide | estimated | 4608x2592 | 100 | 0.15–2.49 | 71 | 0.04 | 0.113 | 38.8 | 94.0 | 11.4 | 62.2 |
| Stereolabs ZED X / X Mini 2.2 mm eye | estimated | 1920x1200 | 110 | 0.15–1.01 | 48 | 0.04 | 0.218 | 41.2 | 94.1 | 12.2 | 64.7 |
| Stereolabs ZED X 4 mm eye | estimated | 1920x1200 | 74 | 0.15–1.99 | 66 | 0.02 | 0.067 | 36.2 | 93.7 | 9.5 | 61.6 |
| Stereolabs ZED 2i 2.1 mm eye | estimated | 2208x1242 | 95 | 0.15–1.78 | 57 | 0.03 | 0.122 | 37.0 | 93.9 | 10.9 | 61.4 |
| Logitech C920 HD Pro Webcam · focus 0.3 m | estimated | 1920x1080 | 70 | – | – | – | – | – | – | 8.4 | – |
On the robot, optics only (the wrist camera's 75 deg view, each camera's sensor and lens)¶
Framing and distortion removed: what each sensor and lens (including its focus) add at the wrist.
| Camera | Status | Image (px) | HFoV (deg) | Detection mAP50 | Segmentation mIoU | Object depth AbsRel, DA-V2 (%) | VLM grounding hit (%) |
|---|---|---|---|---|---|---|---|
| Ideal pinhole (1024x1024, 75 deg) | ideal | 1024x1024 | 75 | 39.5 | 94.3 | 9.1 | 62.8 |
| Intel RealSense D435 / D435i RGB camera · optics only | estimated | 1920x1080 | 108 | 39.8 | 93.9 | 11.6 | 61.1 |
| Intel RealSense D455 RGB camera · optics only | estimated | 1280x800 | 102 | 39.2 | 94.1 | 10.6 | 60.9 |
| Logitech C920 HD Pro Webcam · optics only | estimated | 1920x1080 | 108 | 39.9 | 93.3 | 13.0 | 61.9 |
| Luxonis OAK-D color camera · optics only | estimated | 4056x3040 | 91 | 43.3 | 93.3 | 11.4 | 61.8 |
| Luxonis OAK-D mono camera (one stereo eye) · optics only * | estimated | 1280x800 | 102 | 38.5 | 94.0 | 11.8 | 54.3 |
| Raspberry Pi Camera Module 3 · optics only | estimated | 4608x2592 | 108 | 40.1 | 93.0 | 13.2 | 59.8 |
| Raspberry Pi Camera Module 3 Wide · optics only | estimated | 4608x2592 | 108 | 39.9 | 94.1 | 12.4 | 60.9 |
| Stereolabs ZED X / X Mini 2.2 mm eye · optics only | estimated | 1920x1200 | 102 | 39.9 | 94.2 | 10.9 | 61.3 |
| Stereolabs ZED X 4 mm eye · optics only | estimated | 1920x1200 | 102 | 42.5 | 93.2 | 12.9 | 62.6 |
| Stereolabs ZED 2i 2.1 mm eye · optics only | estimated | 2208x1242 | 108 | 39.7 | 94.0 | 12.0 | 60.6 |
| Logitech C920 HD Pro Webcam · focus 0.3 m · optics only | estimated | 1920x1080 | 108 | – | – | 11.4 | – |
Secondary: from LIBERO's third-person (agent-view) position, as mounted¶
Not where a robot's camera sits; kept because OpenVLA's only camera is this one (the OpenVLA column swaps it).
| Camera | Status | Image (px) | HFoV (deg) | Board range (m) | Board found (%) | Pose error (% of dist.) | Corner error (mrad) | Detection mAP50 | Segmentation mIoU | Object depth AbsRel, DA-V2 (%) | Object depth AbsRel, Depth Pro (%) | VLM grounding hit (%) | OpenVLA success (%) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Ideal pinhole (1024x1024, 45 deg) | ideal | 1024x1024 | 45 | 0.19–1.99 | 66 | 0.03 | 0.068 | 62.7 | 94.7 | 5.3 | 25.0 | 79.0 | 85 (n=100) |
| Intel RealSense D435 / D435i RGB camera | estimated | 1920x1080 | 69 | 0.15–1.99 | 64 | 0.02 | 0.074 | 65.4 | 94.1 | 5.8 | 25.7 | 77.0 | 87 (n=100) |
| Intel RealSense D455 RGB camera | estimated | 1280x800 | 90 | 0.15–0.91 | 44 | 0.06 | 0.240 | 55.2 | 94.0 | 7.0 | 16.4 | 50.1 | 0 (n=100) |
| Intel RealSense D435 / D435i IR imager (one stereo eye) †* | estimated | 1280x800 | 90 | 0.15–1.01 | 45 | 0.05 | 0.220 | 48.4 | 93.4 | 7.1 | 14.9 | 62.9 | – |
| Intel RealSense D455 IR imager (one stereo eye) †* | estimated | 1280x800 | 90 | 0.15–1.01 | 45 | 0.05 | 0.220 | 48.3 | 93.4 | 7.1 | 14.9 | 63.2 | – |
| Logitech C920 HD Pro Webcam | estimated | 1920x1080 | 70 | 0.15–1.78 | 66 | 0.03 | 0.099 | 63.4 | 94.3 | 5.9 | 25.6 | 74.2 | 84 (n=100) |
| Luxonis OAK-D color camera | estimated | 4056x3040 | 69 | 0.15–1.99 | 70 | 0.02 | 0.046 | 63.3 | 94.3 | 5.5 | 24.3 | 61.0 | – |
| Luxonis OAK-D mono camera (one stereo eye) * | estimated | 1280x800 | 72 | 0.15–1.27 | 56 | 0.02 | 0.086 | 55.1 | 94.0 | 5.7 | 20.7 | 70.0 | 65 (n=100) |
| Raspberry Pi Camera Module 3 | estimated | 4608x2592 | 66 | 0.21–3.12 | 84 | 0.05 | 0.106 | 65.8 | 94.4 | 5.6 | 25.0 | 79.2 | – |
| Raspberry Pi Camera Module 3 Wide | estimated | 4608x2592 | 100 | 0.15–2.49 | 71 | 0.04 | 0.113 | 56.3 | 93.3 | 7.5 | 13.9 | 58.8 | – |
| Stereolabs ZED X / X Mini 2.2 mm eye | estimated | 1920x1200 | 110 | 0.15–1.01 | 48 | 0.04 | 0.218 | 54.9 | 93.6 | 6.6 | 15.2 | 44.9 | 0 (n=100) |
| Stereolabs ZED X 4 mm eye | estimated | 1920x1200 | 74 | 0.15–1.99 | 66 | 0.02 | 0.067 | 62.1 | 94.4 | 5.7 | 23.4 | 59.2 | 46 (n=100) |
| Stereolabs ZED 2i 2.1 mm eye | estimated | 2208x1242 | 95 | 0.15–1.78 | 57 | 0.03 | 0.122 | 59.7 | 94.1 | 6.2 | 22.7 | 63.1 | – |
Secondary: third-person position, framed like the reference¶
Each camera moved along its axis to see the same workspace as the reference.
| Camera | Status | Image (px) | HFoV (deg) | Detection mAP50 | Segmentation mIoU | Object depth AbsRel, DA-V2 (%) | Object depth AbsRel, Depth Pro (%) | VLM grounding hit (%) | OpenVLA success (%) |
|---|---|---|---|---|---|---|---|---|---|
| Ideal pinhole (1024x1024, 45 deg) | ideal | 1024x1024 | 45 | 62.7 | 94.7 | 5.3 | 25.0 | 79.0 | 85 (n=100) |
| Intel RealSense D435 / D435i RGB camera · framed | estimated | 1920x1080 | 69 | 62.9 | 94.5 | 5.8 | 25.0 | 73.9 | – |
| Intel RealSense D455 RGB camera · framed | estimated | 1280x800 | 90 | 62.8 | 94.1 | 7.4 | 29.9 | 60.0 | – |
| Logitech C920 HD Pro Webcam · framed | estimated | 1920x1080 | 70 | 62.9 | 94.5 | 5.7 | 24.7 | 71.2 | – |
| Luxonis OAK-D color camera · framed | estimated | 4056x3040 | 69 | 63.2 | 94.1 | 6.0 | 26.3 | 67.8 | – |
| Luxonis OAK-D mono camera (one stereo eye) · framed * | estimated | 1280x800 | 72 | 57.2 | 94.1 | 6.1 | 20.3 | 70.1 | – |
| Raspberry Pi Camera Module 3 · framed | estimated | 4608x2592 | 66 | 64.4 | 94.8 | 5.5 | 21.8 | 71.2 | – |
| Raspberry Pi Camera Module 3 Wide · framed | estimated | 4608x2592 | 100 | 64.8 | 94.0 | 8.7 | 28.7 | 69.3 | – |
| Stereolabs ZED X / X Mini 2.2 mm eye · framed | estimated | 1920x1200 | 110 | 66.3 | 94.1 | 9.6 | 40.3 | 60.8 | – |
| Stereolabs ZED X 4 mm eye · framed | estimated | 1920x1200 | 74 | 64.0 | 94.1 | 5.9 | 26.4 | 62.8 | – |
| Stereolabs ZED 2i 2.1 mm eye · framed | estimated | 2208x1242 | 95 | 64.2 | 93.9 | 6.8 | 34.0 | 71.2 | – |
Secondary: third-person position, optics only¶
LIBERO's 45 deg view with each camera's sensor and lens.
| Camera | Status | Image (px) | HFoV (deg) | Detection mAP50 | Segmentation mIoU | Object depth AbsRel, DA-V2 (%) | Object depth AbsRel, Depth Pro (%) | VLM grounding hit (%) | OpenVLA success (%) |
|---|---|---|---|---|---|---|---|---|---|
| Ideal pinhole (1024x1024, 45 deg) | ideal | 1024x1024 | 45 | 62.7 | 94.7 | 5.3 | 25.0 | 79.0 | 85 (n=100) |
| Intel RealSense D435 / D435i RGB camera · optics only | estimated | 1920x1080 | 73 | 64.1 | 94.3 | 5.9 | 25.2 | 69.8 | – |
| Intel RealSense D455 RGB camera · optics only | estimated | 1280x800 | 67 | 63.6 | 94.0 | 5.6 | 27.5 | 63.0 | 83 (n=48) |
| Logitech C920 HD Pro Webcam · optics only | estimated | 1920x1080 | 73 | 63.9 | 94.3 | 5.9 | 25.1 | 71.4 | 82 (n=40) |
| Luxonis OAK-D color camera · optics only | estimated | 4056x3040 | 58 | 64.7 | 94.3 | 5.1 | 24.6 | 74.8 | – |
| Luxonis OAK-D mono camera (one stereo eye) · optics only * | estimated | 1280x800 | 67 | 57.1 | 94.1 | 5.5 | 21.0 | 71.2 | 85 (n=40) |
| Raspberry Pi Camera Module 3 · optics only | estimated | 4608x2592 | 73 | 64.3 | 94.4 | 5.9 | 23.5 | 71.5 | – |
| Raspberry Pi Camera Module 3 Wide · optics only | estimated | 4608x2592 | 73 | 64.5 | 94.4 | 5.9 | 24.1 | 70.6 | – |
| Stereolabs ZED X / X Mini 2.2 mm eye · optics only | estimated | 1920x1200 | 67 | 65.2 | 94.2 | 5.5 | 26.4 | 66.5 | 85 (n=40) |
| Stereolabs ZED X 4 mm eye · optics only | estimated | 1920x1200 | 67 | 64.5 | 94.2 | 5.6 | 25.7 | 66.8 | – |
| Stereolabs ZED 2i 2.1 mm eye · optics only | estimated | 2208x1242 | 73 | 64.0 | 94.4 | 5.9 | 24.8 | 70.2 | – |
* mono sensor (one channel, shown in all three). † a near-infrared imager rendered in visible light (the simulators render visible light only): its detection, segmentation, depth and grounding scores are not what it would see in the infrared.
How to read it¶
- Simulation only. Every catalog lens is a surrogate (status
estimated): a published glass design scaled to the camera's focal length and f-number, because vendors publish no prescriptions. Sensor noise, colour filter, exposure and motion are not simulated. The table ranks the twins; whether it ranks the real cameras the same way is what the real-capture study is for. - On the robot first. The main tables put each camera where a robot's camera sits: in
place of the robot's wrist camera (LIBERO's
robot0_eye_in_hand, 75 deg vertical), as mounted and optics only (the wrist view's field of view with each camera's sensor and lens). The reference is the wrist pinhole. Depth Pro is left out at the wrist: at 0.1-0.4 m its metric scale fails (errors of 56-111 % for every camera), which says nothing about the cameras. - Secondary: the third-person position. LIBERO's agent-view camera is not on the robot; those tables are kept, marked secondary, because OpenVLA reads only that camera (its column swaps it). The fiducial board is placed relative to each camera, so it is reported once per camera.
- Grounding coordinates: on some image sizes Qwen2.5-VL's boxes are compressed horizontally by 2-3 % (seen on the OAK-D colour, ZED X and D455 colour images; not on others of the same input size). It lowers those rows' grounding hits; it is the model's, not the camera's.
- Native resolution. Images are rendered at each sensor's full resolution. The learned models resize their input (Grounding DINO ~800 px, SAM 1024 px, Depth Anything 518 px, Depth Pro 1536 px, Qwen2.5-VL ~1 MP), so beyond their input size extra pixels do not help them; the fiducial task uses every pixel.
Tasks¶
| Column | Scenes | Model | Metric |
|---|---|---|---|
| Board range, found, pose and corner error | ChArUco board 21 x 15 cm (7 x 5 squares of 30 mm), a range sweep 0.15-12 m and a grid of 36 poses: 6 distances (0.3-3 m) x 3 field positions (centre, 50 % and 80 % of each camera's half field) x 2 tilts (0, 50 deg) | OpenCV ChArUco detection (markers >= 12 px), solvePnP with the camera's calibration | range: nearest-farthest distance with >= 50 % of corners found (frontal, centred); found: corners found over the grid; pose: mean 3D corner error / distance (median); corner: error of found corners vs. their true projection, in mrad (px / focal length) |
| Detection | 200 LIBERO frames: LIBERO-Spatial demo_0 and demo_1 of all 10 tasks at 0, 1/3 and 2/3 of the episode (60); LIBERO-Object init states 0-3 (40); LIBERO-Goal and LIBERO-10 init states 0-4 (50 each) | Grounding DINO base, prompted with the scene's object phrases | AP at IoU 0.5 per phrase, mean over phrases |
| Segmentation | same frames | SAM ViT-H, prompted with each object's true box | mask IoU with the true visible mask |
| Object depth | same frames | Depth Anything V2 Large (relative: scale and shift fitted in inverse depth over the whole image); Depth Pro (metric, given the camera's calibrated focal length) | AbsRel on the target objects' pixels (whole-image errors are in the cells: they depend on how much floor and wall a camera's view holds) |
| VLM grounding | same frames | Qwen2.5-VL-7B-Instruct, greedy, ~1 MP input: "Locate the {object} ..." | hit: the returned box's centre lies on the object |
| OpenVLA success | LIBERO-Spatial, 10 tasks x 10 initial states | OpenVLA-7B fine-tuned on LIBERO-Spatial (closed loop) | task success; only the cameras run there |
Targets are the objects with at least 16 visible pixels in a camera's image.
95 % intervals¶
Bootstrap over images (fixed seed); OpenVLA: Wilson interval over episodes.
| Camera | Board found (%) | Detection mAP50 | Segmentation mIoU | Object depth AbsRel, DA-V2 (%) | Object depth AbsRel, Depth Pro (%) | VLM grounding hit (%) | OpenVLA success (%) |
|---|---|---|---|---|---|---|---|
| Ideal pinhole (1024x1024, 45 deg) | 52.5–79.6 | 60.1–65.7 | 94.3–95.1 | 5.0–5.6 | 22.3–27.6 | 77.2–80.8 | 76.7–90.7 |
| Intel RealSense D435 / D435i RGB camera | 49.5–77.8 | 63.4–67.9 | 93.7–94.5 | 5.5–6.1 | 23.4–28.0 | 74.6–79.6 | 79.0–92.2 |
| Intel RealSense D455 RGB camera | 28.9–58.5 | 53.5–57.5 | 93.5–94.4 | 6.6–7.4 | 14.9–18.0 | 47.5–52.8 | 0.0–3.7 |
| Intel RealSense D435 / D435i IR imager (one stereo eye) | 30.0–60.1 | 46.4–51.0 | 92.9–93.9 | 6.7–7.5 | 13.2–16.8 | 60.4–65.3 | – |
| Intel RealSense D455 IR imager (one stereo eye) | 30.0–60.1 | 46.4–51.0 | 92.9–93.9 | 6.7–7.5 | 13.2–16.8 | 60.7–65.6 | – |
| Logitech C920 HD Pro Webcam | 51.5–79.3 | 61.5–66.2 | 93.9–94.7 | 5.5–6.2 | 23.2–28.0 | 72.4–76.2 | 75.6–89.9 |
| Luxonis OAK-D color camera | 58.9–81.7 | 61.7–65.9 | 93.9–94.6 | 5.1–5.8 | 21.7–26.8 | 58.3–63.7 | – |
| Luxonis OAK-D mono camera (one stereo eye) | 40.9–70.7 | 53.1–57.4 | 93.6–94.4 | 5.4–6.0 | 18.3–23.1 | 67.7–72.1 | 55.3–73.6 |
| Raspberry Pi Camera Module 3 | 75.3–92.5 | 64.1–68.2 | 94.0–94.9 | 5.3–5.9 | 22.7–27.3 | 77.0–81.3 | – |
| Raspberry Pi Camera Module 3 Wide | 58.0–82.8 | 54.8–58.7 | 92.8–93.8 | 7.0–7.9 | 12.7–15.0 | 55.8–61.6 | – |
| Stereolabs ZED X / X Mini 2.2 mm eye | 33.4–62.4 | 52.9–57.1 | 93.1–94.0 | 6.1–7.0 | 14.1–16.2 | 42.3–47.7 | 0.0–3.7 |
| Stereolabs ZED X 4 mm eye | 52.5–79.2 | 60.2–64.8 | 94.0–94.8 | 5.4–6.1 | 21.0–25.7 | 56.4–61.9 | 36.6–55.7 |
| Stereolabs ZED 2i 2.1 mm eye | 42.5–71.5 | 58.1–61.9 | 93.7–94.5 | 5.8–6.6 | 20.3–25.0 | 60.5–65.7 | – |
| Intel RealSense D435 / D435i RGB camera · framed | – | 61.3–65.3 | 94.1–94.9 | 5.5–6.2 | 22.5–27.3 | 72.0–76.0 | – |
| Intel RealSense D455 RGB camera · framed | – | 61.0–65.5 | 93.7–94.5 | 6.9–7.8 | 26.8–33.1 | 57.0–62.9 | – |
| Logitech C920 HD Pro Webcam · framed | – | 61.0–65.4 | 94.2–94.9 | 5.4–6.1 | 22.2–27.0 | 68.7–73.6 | – |
| Luxonis OAK-D color camera · framed | – | 60.2–66.2 | 93.7–94.6 | 5.6–6.3 | 23.3–29.3 | 65.4–70.4 | – |
| Luxonis OAK-D mono camera (one stereo eye) · framed | – | 55.4–59.3 | 93.6–94.5 | 5.7–6.4 | 18.0–22.6 | 67.8–72.4 | – |
| Raspberry Pi Camera Module 3 · framed | – | 62.7–66.8 | 94.4–95.1 | 5.2–5.8 | 19.5–24.0 | 69.3–73.4 | – |
| Raspberry Pi Camera Module 3 Wide · framed | – | 63.1–67.2 | 93.6–94.4 | 8.2–9.3 | 26.0–31.5 | 66.9–71.8 | – |
| Stereolabs ZED X / X Mini 2.2 mm eye · framed | – | 64.7–69.0 | 93.8–94.6 | 9.0–10.1 | 37.1–43.5 | 58.4–63.4 | – |
| Stereolabs ZED X 4 mm eye · framed | – | 62.1–67.0 | 93.6–94.6 | 5.6–6.3 | 23.7–29.1 | 60.2–65.5 | – |
| Stereolabs ZED 2i 2.1 mm eye · framed | – | 62.5–66.7 | 93.5–94.3 | 6.4–7.3 | 30.6–37.6 | 68.9–73.6 | – |
| Intel RealSense D435 / D435i RGB camera · optics only | – | 62.2–66.5 | 93.9–94.7 | 5.6–6.2 | 22.7–27.7 | 67.3–72.1 | – |
| Intel RealSense D455 RGB camera · optics only | – | 61.5–66.2 | 93.6–94.4 | 5.2–5.9 | 25.0–30.1 | 59.9–65.9 | 70.4–91.3 |
| Logitech C920 HD Pro Webcam · optics only | – | 62.2–66.4 | 93.9–94.7 | 5.6–6.2 | 22.6–27.6 | 69.2–73.5 | 68.0–91.3 |
| Luxonis OAK-D color camera · optics only | – | 62.9–67.3 | 93.8–94.7 | 4.8–5.3 | 22.0–27.1 | 72.7–77.0 | – |
| Luxonis OAK-D mono camera (one stereo eye) · optics only | – | 55.3–59.7 | 93.6–94.5 | 5.2–5.8 | 18.7–23.4 | 68.9–73.3 | 70.9–92.9 |
| Raspberry Pi Camera Module 3 · optics only | – | 62.6–66.5 | 94.1–94.8 | 5.6–6.2 | 21.2–25.7 | 69.3–73.6 | – |
| Raspberry Pi Camera Module 3 Wide · optics only | – | 62.6–66.8 | 94.0–94.8 | 5.6–6.2 | 21.6–26.4 | 68.0–73.0 | – |
| Stereolabs ZED X / X Mini 2.2 mm eye · optics only | – | 63.1–68.0 | 93.7–94.6 | 5.2–5.9 | 23.9–28.9 | 63.7–69.2 | 70.9–92.9 |
| Stereolabs ZED X 4 mm eye · optics only | – | 62.6–67.0 | 93.8–94.7 | 5.2–5.9 | 23.3–28.1 | 64.0–69.6 | – |
| Stereolabs ZED 2i 2.1 mm eye · optics only | – | 62.3–66.4 | 94.0–94.8 | 5.6–6.2 | 22.3–27.2 | 67.6–72.5 | – |
| Ideal pinhole (1024x1024, 75 deg) | – | 37.5–43.8 | 93.7–94.9 | 8.4–9.9 | 76.6–99.9 | 60.3–65.2 | 76.7–90.7 |
| Intel RealSense D435 / D435i RGB camera | 49.5–77.8 | 31.6–38.3 | 92.5–93.8 | 7.1–9.8 | 77.6–100.0 | 56.8–62.5 | 79.0–92.2 |
| Intel RealSense D455 RGB camera | 28.9–58.5 | 33.4–40.2 | 93.4–94.5 | 8.8–10.7 | 70.8–87.3 | 58.5–64.0 | 0.0–3.7 |
| Intel RealSense D435 / D435i IR imager (one stereo eye) | 30.0–60.1 | 31.3–38.0 | 93.8–94.9 | 8.3–10.9 | 49.7–62.3 | 50.0–55.7 | – |
| Intel RealSense D455 IR imager (one stereo eye) | 30.0–60.1 | 31.3–38.0 | 93.8–94.9 | 8.3–10.9 | 49.7–62.3 | 49.8–55.6 | – |
| Logitech C920 HD Pro Webcam | 51.5–79.3 | 33.3–40.5 | 91.9–93.4 | 7.5–9.8 | 85.9–106.6 | 59.2–65.0 | 75.6–89.9 |
| Luxonis OAK-D color camera | 58.9–81.7 | 35.2–42.4 | 92.5–93.8 | 8.5–11.0 | 98.5–122.8 | 58.2–63.4 | – |
| Luxonis OAK-D mono camera (one stereo eye) | 40.9–70.7 | 28.0–34.7 | 93.2–94.5 | 8.1–10.5 | 78.6–100.8 | 48.5–54.5 | 55.3–73.6 |
| Raspberry Pi Camera Module 3 | 75.3–92.5 | 32.9–40.0 | 92.0–93.4 | 7.8–10.0 | 86.6–108.7 | 54.1–59.7 | – |
| Raspberry Pi Camera Module 3 Wide | 58.0–82.8 | 36.6–42.6 | 93.4–94.5 | 9.9–13.0 | 69.0–84.2 | 59.2–65.0 | – |
| Stereolabs ZED X / X Mini 2.2 mm eye | 33.4–62.4 | 39.2–44.7 | 93.7–94.6 | 11.0–13.3 | 90.1–105.7 | 62.3–67.2 | 0.0–3.7 |
| Stereolabs ZED X 4 mm eye | 52.5–79.2 | 33.5–41.3 | 93.0–94.3 | 8.3–10.7 | 86.7–106.8 | 59.0–64.3 | 36.6–55.7 |
| Stereolabs ZED 2i 2.1 mm eye | 42.5–71.5 | 34.7–41.0 | 93.4–94.4 | 9.5–12.6 | 95.2–117.7 | 58.4–64.2 | – |
| Intel RealSense D435 / D435i RGB camera · optics only | – | 37.8–43.4 | 93.4–94.4 | 10.5–12.8 | 63.1–75.9 | 58.2–63.7 | – |
| Intel RealSense D455 RGB camera · optics only | – | 37.1–43.1 | 93.6–94.6 | 9.6–11.5 | 64.2–78.0 | 58.4–63.2 | 70.4–91.3 |
| Logitech C920 HD Pro Webcam · optics only | – | 38.0–43.8 | 92.7–93.8 | 11.7–14.4 | 62.7–75.7 | 59.4–64.4 | 68.0–91.3 |
| Luxonis OAK-D color camera · optics only | – | 41.4–47.3 | 92.8–93.7 | 10.2–12.7 | 80.7–102.3 | 59.2–64.6 | – |
| Luxonis OAK-D mono camera (one stereo eye) · optics only | – | 36.2–41.5 | 93.5–94.5 | 10.5–13.2 | 50.4–60.5 | 51.4–56.8 | 70.9–92.9 |
| Raspberry Pi Camera Module 3 · optics only | – | 38.3–44.0 | 92.4–93.5 | 11.6–14.7 | 68.6–83.7 | 57.2–62.2 | – |
| Raspberry Pi Camera Module 3 Wide · optics only | – | 37.8–43.7 | 93.6–94.6 | 11.1–13.8 | 63.1–76.6 | 58.5–63.5 | – |
| Stereolabs ZED X / X Mini 2.2 mm eye · optics only | – | 37.9–43.7 | 93.7–94.7 | 9.8–11.9 | 64.2–77.9 | 58.8–63.7 | 70.9–92.9 |
| Stereolabs ZED X 4 mm eye · optics only | – | 40.2–45.8 | 92.7–93.7 | 11.6–14.3 | 69.4–85.1 | 59.8–65.3 | – |
| Stereolabs ZED 2i 2.1 mm eye · optics only | – | 37.5–43.5 | 93.5–94.5 | 10.8–13.2 | 63.2–76.0 | 58.2–63.0 | – |
| Logitech C920 HD Pro Webcam · focus 0.3 m | – | – | – | 7.1–9.9 | 76.9–96.4 | – | – |
| Logitech C920 HD Pro Webcam · focus 0.3 m · optics only | – | – | – | 10.3–12.5 | 59.7–72.1 | – | – |
Provenance¶
- TwinRobo-Core: 075020f, 48bad68
- da2:
depth-anything/Depth-Anything-V2-Large-hf@7581137eff8d4e94f6e796d3baea0e9fa79b22d2 - depthpro:
apple/DepthPro-hf@de816c8ce7168afcb231f96d501d72b869d0beda - gdino:
IDEA-Research/grounding-dino-base@12bdfa3120f3e7ec7b434d90674b3396eccf88eb - policy:
openvla/openvla-7b-finetuned-libero-spatial@962318cec55ac10993ff0f5f43eda9a270b4c873 - qwen:
Qwen/Qwen2.5-VL-7B-Instruct@cc594898137f460bfe9f0759e9844b3ce807cfb5 - sam:
facebook/sam-vit-huge@87aecf0df4ce6b30cd7de76e87673c49644bdf67 - software: deeplens-core (render) 2.5.4, mujoco (render) 3.3.2, numpy (models) 2.5.2, numpy (render) 2.5.3, opencv-python (render) 5.0.0.93, opencv-python-headless (models) 5.0.0.93, torch (models) 2.8.0+cu128, torch (render) 2.10.0, transformers (models) 5.19.0