Ouster OS0-128: Spinning Lidar Benchmark

• Ouster OS Series is a group of high-end spinning lidars, exemplified by the OS0-128, featuring uniform 360° horizontal and 90° vertical fields of view with integrated IMU.
• It is employed as a precise reference in comparative studies, enabling controlled evaluation of SLAM and ICP performance against solid-state sensors.
• The sensor’s integration on a calibrated multi-sensor platform demonstrates actionable insights for range tuning, synchronization, and feature extraction in various environments.

Searching arXiv for the cited paper and related lidar dataset work to ground the article in current literature. {"query":"arXiv (Felix et al., 6 Jul 2025) Lidar Variability: A Novel Dataset and Comparative Study of Solid-State and Spinning Lidars", "max_results": 5} The Ouster OS Series, in the comparative framework examined by "Lidar Variability: A Novel Dataset and Comparative Study of Solid-State and Spinning Lidars," denotes the high-end spinning lidar family represented experimentally by the Ouster OS0-128, a spinning multi-beam sensor used as the principal contrast class against the low-cost solid-state Livox Avia and Livox Mid-360 (Felix et al., 6 Jul 2025). In that study, the OS0-128 is not treated as a generic product-line abstraction but as a concrete sensing modality with regular angular sampling, a $360^\circ$ horizontal field of view, a nominal $90^\circ$ vertical field of view, and an integrated IMU. Its significance lies in enabling a controlled comparison between spinning and solid-state lidars on a single calibrated robotic platform, under both lidar-inertial SLAM and IMU-free ICP registration.

1. Sensor identity and role in the comparative study

Within the reported experiments, the only Ouster device actually used is the Ouster OS0-128; no OS1 or OS2 unit is used in the dataset or benchmarks, although OS-series variants appear in the discussion of earlier datasets (Felix et al., 6 Jul 2025). The paper explicitly positions the OS0-128 as the representative high-end spinning lidar, contrasted with two low-cost solid-state Livox sensors.

From Table 2, the OS0-128 is specified as a spinning, multi-beam lidar with $360^\circ$ horizontal FoV, $90^\circ$ nominal vertical FoV, and max range "Up to 240 m". Its hardware specifications are reported as 14–20 W (28 W peak) power, 10–51 V supply voltage, 430–500 g weight, and an integrated IAM-20680HT IMU. In the dataset and experiments, it is used with a 10 Hz point cloud stream, designated "Ouster_10Hz" in the benchmark tables.

Property	OS0-128 value	Experimental usage
Type	Spinning, multi-beam	High-end spinning reference
FoV	$360^\circ$ horizontal; $90^\circ$ nominal vertical	Compared directly with Avia and Mid-360
Max range	Up to 240 m	Truncated to 60 m indoors; 150 m in outdoor SLAM
IMU	IAM-20680HT	Available via /ouster/imu at 100 Hz

The experimental range policy is operationally important. Indoors, the point cloud range is truncated to 60 m for all lidars to standardize evaluation. In outdoor SLAM, the OS0 is used with a 150 m maximum range, versus 450 m for Avia and 100 m for Mid-360. For outdoor ICP subsegments, 60 m is used uniformly across all sensors. These settings materially shape the comparative outcomes, particularly in long-road outdoor trajectories where available range influences constraint accumulation.

2. Platform integration, calibration, and synchronized data products

All data are collected on a Unitree B1 quadruped robot carrying a compact multi-sensor rack comprising the Ouster OS0-128, Livox Avia, Livox Mid-360, Intel RealSense L515, Xsens MTI-680G, and a GNSS-RTK antenna (Felix et al., 6 Jul 2025). The lidars are rigidly mounted on a common calibrated frame. Although the exact mounting height and orientation are not numerically listed, the paper states that the platform is calibrated, implying extrinsic calibration among the lidars, the IMU, and the reference frames used for evaluation.

Precision Time Protocol (PTP) is used to synchronize all sensors, including the Ouster OS0. This synchronization is central to the dataset design because the three lidars and the Xsens unit are recorded simultaneously in linked ROS bags, permitting cross-sensor comparisons under matched motion and scene conditions.

The Ouster point cloud is recorded in ROS sensor_msgs/PointCloud2 format. The paper specifies the per-point fields as x, y, z, intensity, reflectivity, ring, ambient, range, and t, where t is a timestamp offset in nanoseconds relative to the scan. These attributes are relevant for motion correction, channel-wise modeling, and intensity or reflectivity-aware processing. The associated Ouster IMU topic is /ouster/imu, with message type sensor_msgs/Imu and frequency 100 Hz.

For ground truth, the indoor sequences use MoCap via /vrpn_client_node/unitree_b1/pose, while the outdoor sequence uses the Xsens MTI-680G GNSS-RTK topic /gnss_pose. The Xsens system is described as providing IMU + GNSS-RTK, 100 Hz, cm-level reference outdoors. ICP experiments are explicitly IMU-free, using only geometry and timestamps; this isolates lidar-type effects from inertial aiding.

3. Dataset coverage and sequence structure

The dataset contains three principal sequences, all including Ouster OS0-128 data (Felix et al., 6 Jul 2025). The first two, IndoorOffice1 and IndoorOffice2, are recorded in a structured indoor environment described as "factory-like" or office. Both are evaluated against MoCap ground truth and are designed to test trajectory-following under confined geometry, turns, and planar surfaces. The third sequence, OutdoorRoad, is an unstructured/mixed outdoor "open-road" environment containing natural and urban elements and evaluated against the Xsens GNSS-RTK solution.

For local registration, the paper further defines OutdoorRoad-cut0 and OutdoorRoad-cut1, shorter subsegments of the outdoor trajectory selected because pure ICP lacks loop closure and would otherwise suffer severe drift. This segmentation is methodologically significant: it constrains evaluation to local registration fidelity rather than long-horizon global consistency.

The paper situates this dataset against prior multi-lidar corpora. It notes that the TIERS Multi-Modal Lidar Dataset included Ouster OS0-128 / OS1-64 but not Mid-360; GEODE included Ouster OS1-32 and Livox Avia but not dome sensors and not IMU-free registration; and CTE-MLO included OS1-64 and Mid-360, but with Mid-360 on a MAV rather than on the same platform as Ouster. The contribution claimed here is therefore a joint dataset in which Ouster OS0-128, Livox Avia, and Livox Mid-360 are rigidly mounted on the same robot, operated in the same environment at the same time, and benchmarked under both SLAM and ICP.

4. SLAM performance of the OS0-128

The SLAM benchmark comprises five lidar-inertial algorithms: FAST-LIO2, FASTER-LIO, S-FAST-LIO, GLIM, and FAST-LIO-SAM (Felix et al., 6 Jul 2025). Each algorithm is run independently on Avia_10Hz, Mid360_10Hz, and Ouster_10Hz. Indoors, all sensors are limited to 60 m; outdoors, default working ranges are used, namely Avia 450 m, Mid-360 100 m, and Ouster 150 m.

In the indoor sequences, the OS0-128 performs strongly but is generally marginally worse than Mid-360. For FAST-LIO2 on IndoorOffice1, the APE is 0.0446 ± 0.0298 m for Ouster, versus 0.0451 ± 0.0150 m for Mid-360 and 0.1436 ± 0.1390 m for Avia. For FASTER-LIO on IndoorOffice2, Ouster records 0.0549 ± 0.0491 m, compared with 0.0460 ± 0.0147 m for Mid-360 and 0.2692 ± 0.1386 m for Avia. For S-FAST-LIO on IndoorOffice1, Ouster yields 0.0541 ± 0.0367 m, while Mid-360 yields 0.0427 ± 0.0161 m and Avia 0.1808 ± 0.1840 m. The paper’s synthesis is explicit: Mid-360 consistently yields the best indoor SLAM performance, and Ouster OS0-128 is very close, usually second best and often within approximately 1 cm of Mid-360 in mean APE.

The outdoor results differ. On OutdoorRoad, Ouster is not dominant and may be mid-pack or worst, depending on the algorithm. With FAST-LIO2, Ouster records 0.5845 ± 0.3127 m, compared with 0.3755 ± 0.1527 m for Avia and 0.3893 ± 0.1788 m for Mid-360. With FASTER-LIO, Ouster reaches 0.4245 ± 0.2273 m, while Avia achieves the best reported value, 0.3013 ± 0.0818 m; Mid-360 obtains 0.3666 ± 0.1668 m. With S-FAST-LIO, Ouster gives 0.6223 ± 0.3282 m, versus 0.6730 ± 0.3205 m for Avia and 0.3721 ± 0.1641 m for Mid-360.

The interpretation offered by the paper is that Ouster is competitive but not superior. Indoors, the Mid-360 dome geometry provides broader vertical coverage and thus stronger floor-and-ceiling constraints. Outdoors, particularly on long, mostly planar road segments, range can matter more than vertical coverage, so Avia’s 450 m range can benefit methods such as FASTER-LIO. The authors’ final summary is that "Ouster sensors achieved competitive results, especially under spinning-optimized SLAM pipelines, though their limited vertical resolution reduced performance in semi-structured environments."

5. ICP-based registration and IMU-free odometry behavior

The ICP study is explicitly designed to remove inertial assistance and expose the effect of lidar characteristics alone (Felix et al., 6 Jul 2025). Three registration families are examined: point-to-point, point-to-plane, and hybrid methods, instantiated respectively by KISS-ICP, Open3D-GICP, and GenZ-ICP.

The point-to-point objective is given as

$$\min_{R,\mathbf{t}} \sum_{i=1}^{N} \left\| \mathbf{p}_i - (R \mathbf{q}_i + \mathbf{t}) \right\|^2.$$

The point-to-plane objective is reported as

$$\min_{R,\mathbf{t}} \sum_{i=1}^{N} \left( \mathbf{n}_i^\top (R \mathbf{p}_i + \mathbf{t} - \mathbf{q}_i) \right)^2,$$

and also discussed in the equivalent projected form

$$\sum_{i=1}^{N} \left\| \mathbf{p}_i - (\mathbf{q}_i + \lambda \mathbf{n}_i) \right\|^2.$$

The hybrid formulation is

$$\min_{R,\mathbf{t}} \ \alpha \sum_{i=1}^{N} \left\| \mathbf{p}_i - (R \mathbf{q}_i + \mathbf{t}) \right\|^2 + (1 - \alpha) \sum_{i=1}^{N} \left( \mathbf{n}_i^\top (R \mathbf{p}_i + \mathbf{t} - \mathbf{q}_i) \right)^2,$$

with $90^\circ$0 balancing the two terms.

In indoor ICP, Ouster is generally solid but again not the best overall. Under KISS-ICP, IndoorOffice1 yields 0.1042 ± 0.0838 m for Ouster, against 0.0483 ± 0.0405 m for Mid-360 and 0.1348 ± 0.1049 m for Avia. On IndoorOffice2, Ouster yields 0.0996 ± 0.0782 m, while Mid-360 records 0.0441 ± 0.0286 m and Avia deteriorates to 0.6945 ± 0.2825 m. Under GenZ-ICP, Ouster attains 0.0937 ± 0.0511 m on IndoorOffice1 and 0.1037 ± 0.0764 m on IndoorOffice2; the latter is better than Mid-360’s 0.1214 ± 0.0586 m for that specific setting, but the overall indoor pattern still favors Mid-360. Under Open3D-GICP (Scan2Scan), Avia can fail badly with errors above 1 m, whereas Ouster and Mid-360 remain around 0.1–0.22 m.

Outdoor ICP uses a common 60 m range cap for fairness. Under KISS-ICP on OutdoorRoad-cut0, Ouster achieves 0.0787 ± 0.0502 m, compared with 0.0545 ± 0.0424 m for Mid-360 and 0.3917 ± 0.3135 m for Avia. On cut1, Ouster records 0.1072 ± 0.0559 m, almost identical to Mid-360’s 0.1058 ± 0.0628 m and clearly better than Avia’s 0.2840 ± 0.2149 m. Under GenZ-ICP, Ouster remains slightly worse than Mid-360 but still reasonable. Under Open3D-GICP (Scan2Map), Ouster is best on cut1 with 0.0803 ± 0.0549 m, versus 0.0953 ± 0.0513 m for Mid-360.

The principal outcome is that Mid-360 is overall best for ICP, with the lowest errors and smallest variance across many settings, while Ouster OS0 performs consistently reasonably well and can occasionally be best in specific scan-to-map outdoor segments. A common assumption that a high-end spinning lidar necessarily dominates low-cost solid-state sensors is therefore not supported by these IMU-free experiments.

6. Sensor characteristics, methodological implications, and practical interpretation

The paper does not construct explicit analytic noise models for Ouster, but it does identify systematic behavioral consequences of sensor geometry (Felix et al., 6 Jul 2025). The OS0-128 offers a uniform horizontal angular sampling over $90^\circ$1 with fixed vertical channels over roughly $90^\circ$2, a structure that matches the assumptions of many spinning-lidar SLAM pipelines. Its regularity, together with per-point attributes such as ring, reflectivity, ambient, and t, makes it well aligned with classical feature extraction, motion compensation, and scan organization strategies.

At the same time, the study highlights specific limitations. Relative to Mid-360, the OS0-128 provides less vertical FoV and fewer overhead and near-ground returns in some poses. In semi-structured indoor environments, the paper argues that vertical coverage and density can be more important than complete horizontal coverage, which helps explain why the dome-shaped Mid-360 frequently outperforms Ouster in both SLAM and ICP. Outdoors, Ouster’s 150 m experimental limit can be disadvantageous on long, straight segments when compared with Avia’s 450 m range.

Several practical implications follow directly from the results. Algorithms assuming regular spinning patterns, including the FAST-LIO family and GLIM, are a natural fit for Ouster data. However, spinning geometry alone does not ensure superior performance; the spatial distribution of returns across ground, ceiling, and side-wall structures remains decisive for observability and convergence. The paper further indicates that scan-to-map registration may be more stable than scan-to-scan in unstructured scenes for Ouster, and that careful max-range tuning matters: for local ICP, 60 m is sufficient and improves convergence stability.

The dataset’s intended utility is correspondingly methodological. Once released, it is expected to provide ROS bag files with Ouster point clouds and IMU, aligned ground truth, and calibration information. The study recommends clipping range according to application, using ring and t for motion compensation or per-ring calibration, and, for fair comparison with Livox devices, equalizing either range or point density. In this sense, the Ouster OS Series—as instantiated here by the OS0-128—serves less as a nominal benchmark winner than as a precisely characterized spinning-lidar baseline against which heterogeneous lidar modalities can be compared under tightly synchronized conditions.