--- license: cc-by-4.0 pretty_name: MVSim-LIO tags: - robotics - lidar - imu - lidar-inertial-odometry - lidar-odometry - slam - point-cloud - simulation - semantic-labels - mrpt task_categories: - robotics size_categories: - n<1K viewer: false --- # MVSim-LIO: a simulated LiDAR-inertial benchmark with exact ground truth ![Exact ground-truth maps and trajectories of six of the worlds](https://huggingface.co/datasets/MOLAorg/MVSim-LIO/resolve/main/assets/banner.gif) Exact maps (noise-free sweeps at the true poses) revealed along the ground-truth trajectory; inset: the MVSim world. 15 [MVSim](https://github.com/MRPT/mvsim) worlds and 30 predefined 6-DoF trajectories (20.1 km, 148 minutes) of five carriers (car, UGV, legged, hand-held, quadruped dog), rendered as MRPT `.rawlog` datasets: a rotating 3D LiDAR with per-point timestamps, a 200 Hz IMU and wheel odometry, with exact ground truth. Each trajectory is rendered with nine sensor variants (270 sequences, 266 GB). - **Exact ground truth**: the trajectory is the simulator's input (a C2 spline you can evaluate at any time), and the map is the world's analytic geometry. No GNSS/INS, total station or TLS registration error, no extrinsic calibration or time synchronization error. - **Paired variants**: every trajectory is rendered with global and rolling shutter, and with three noise levels (noise-free, Gaussian, and a realistic LiDAR error model), sharing the same random seeds, so that each factor can be studied on its own. - **Realistic sensor errors**: range noise growing with range, grazing-angle dropouts, beam divergence, mixed pixels at depth edges and per-beam calibration errors, with parameters drawn per sequence; MEMS-grade IMU noise with bias random walk. - **Per-point semantic and instance labels**. Uses: evaluation of LiDAR and LiDAR-inertial odometry and SLAM, on scenes that range from well constrained (streets, campus) to degenerate (tunnel, open field, greenhouse aisles), with controlled sensor noise and motion distortion; and training data for learning-based methods, with ground truth free of any error. ## Download With the [Hugging Face CLI](https://huggingface.co/docs/huggingface_hub/guides/cli) (`hf`): ```bash # Everything (266 GB): hf download MOLAorg/MVSim-LIO --repo-type dataset --local-dir mvsim_lio # One variant, with the 200 Hz ground truth of every sequence: hf download MOLAorg/MVSim-LIO --repo-type dataset --local-dir mvsim_lio \ --include "*/*/os1_32-real-rolling.*" --include "*/*/gt_200hz.tum" # One sequence, all its variants: hf download MOLAorg/MVSim-LIO --repo-type dataset --local-dir mvsim_lio --include "office/spins/*" ``` Size of each variant (all 30 sequences): | variant | size | variant | size | |---|---|---|---| | `os1_32-clean-global` | 21.5 GB | `os1_32-real-rolling` | 24.6 GB | | `os1_32-clean-rolling` | 22.1 GB | `vlp16-noisy-rolling` | 24.0 GB | | `os1_32-noisy-global` | 25.5 GB | `os1_64-noisy-rolling` | 50.6 GB | | `os1_32-noisy-rolling` | 25.9 GB | `os1_64-real-rolling` | 47.6 GB | | `os1_32-real-global` | 24.1 GB | `gt_200hz.tum` (shared) | 0.2 GB | Each file's size and SHA-256 checksum are in `MANIFEST.tsv` and `SHA256SUMS` (`sha256sum -c SHA256SUMS` in the download directory; with a partial download, check only the lines of the files you have). ## Sequences Every trajectory starts and ends with 3 s at rest (for IMU initialization). | sequence | carrier | duration | length | max speed | max yaw rate | max pitch / roll | z range | |---|---|---|---|---|---|---|---| | `apartment_building/dog_floors` | dog | 7:32 | 273 m | 1.4 m/s | 118 deg/s | 29 / 2 deg | 6.0 m | | `apartment_building/dog_stairs` | dog | 3:21 | 121 m | 1.6 m/s | 122 deg/s | 29 / 2 deg | 6.0 m | | `campus/long_tour` | ugv | 15:13 | 1796 m | 2.0 m/s | 91 deg/s | 2 / 2 deg | 5.6 m | | `campus/ring_drive` | car | 6:19 | 2792 m | 8.0 m/s | 47 deg/s | 2 / 2 deg | 4.2 m | | `forest/trail_loop` | legged | 5:40 | 425 m | 1.5 m/s | 22 deg/s | 12 / 11 deg | 8.0 m | | `forest/crossing` | ugv | 2:17 | 195 m | 1.5 m/s | 5 deg/s | 11 / 10 deg | 6.4 m | | `greenhouse/aisles` | ugv | 6:07 | 414 m | 1.2 m/s | 82 deg/s | 0 / 0 deg | 0.0 m | | `greenhouse/walk` | handheld | 4:14 | 265 m | 1.2 m/s | 93 deg/s | 1 / 1 deg | 0.0 m | | `office/walk` | handheld | 3:53 | 226 m | 1.3 m/s | 96 deg/s | 1 / 2 deg | 0.1 m | | `office/spins` | ugv | 2:39 | 120 m | 1.0 m/s | 179 deg/s | 0 / 0 deg | 0.0 m | | `open_field/lap` | ugv | 6:33 | 769 m | 2.0 m/s | 66 deg/s | 5 / 4 deg | 7.2 m | | `open_field/zigzag` | ugv | 4:42 | 702 m | 3.0 m/s | 169 deg/s | 5 / 5 deg | 7.4 m | | `orchard/rows` | ugv | 7:53 | 696 m | 1.5 m/s | 82 deg/s | 3 / 3 deg | 4.5 m | | `orchard/headland` | ugv | 3:41 | 525 m | 2.5 m/s | 82 deg/s | 2 / 3 deg | 5.5 m | | `parking_garage/walk_down` | handheld | 4:01 | 315 m | 1.5 m/s | 96 deg/s | 10 / 2 deg | 6.5 m | | `parking_garage/up_down` | car | 2:51 | 565 m | 3.5 m/s | 74 deg/s | 9 / 0 deg | 6.4 m | | `parking_lot/search` | car | 3:42 | 934 m | 4.5 m/s | 47 deg/s | 0 / 0 deg | 0.0 m | | `parking_lot/ugv_loop` | ugv | 5:10 | 569 m | 2.0 m/s | 143 deg/s | 0 / 0 deg | 0.0 m | | `rough_terrain/hill_loop` | ugv | 6:58 | 623 m | 1.6 m/s | 72 deg/s | 22 / 19 deg | 14.1 m | | `rough_terrain/traverse_fast` | ugv | 1:57 | 329 m | 3.2 m/s | 35 deg/s | 22 / 13 deg | 11.4 m | | `stair_course/dog_course` | dog | 2:40 | 110 m | 1.5 m/s | 134 deg/s | 30 / 2 deg | 6.0 m | | `stairwell/dog_floors` | dog | 3:49 | 163 m | 1.4 m/s | 100 deg/s | 30 / 2 deg | 7.0 m | | `stairwell/legged_floors` | legged | 4:48 | 207 m | 1.0 m/s | 87 deg/s | 32 / 1 deg | 7.0 m | | `stairwell/handheld_stairs` | handheld | 2:40 | 118 m | 1.2 m/s | 85 deg/s | 7 / 1 deg | 3.5 m | | `tunnel/car_through` | car | 2:28 | 1782 m | 15.0 m/s | 32 deg/s | 0 / 0 deg | 0.0 m | | `tunnel/ugv_inspection` | ugv | 5:28 | 798 m | 2.5 m/s | 2 deg/s | 0 / 0 deg | 0.0 m | | `urban_street/car_loops` | car | 5:48 | 2935 m | 10.0 m/s | 47 deg/s | 0 / 0 deg | 0.0 m | | `urban_street/pedestrian` | handheld | 8:23 | 671 m | 1.5 m/s | 90 deg/s | 5 / 2 deg | 0.2 m | | `warehouse/aisles` | ugv | 3:38 | 410 m | 2.0 m/s | 82 deg/s | 0 / 0 deg | 0.0 m | | `warehouse/walk` | handheld | 3:20 | 260 m | 1.4 m/s | 88 deg/s | 1 / 2 deg | 0.1 m | The worlds: an apartment building and a stairwell building (stairs, identical floors), a university campus, a forest on hills, a multi-span greenhouse, an office floor (a long corridor), an open field with few features, an orchard, a three-level parking garage, an open-air parking lot, rough terrain with boulders, a courtyard with straight and spiral stairs, a 735 m road tunnel, a city grid, and a logistics warehouse with pallet racks. ## Sensors and variants Files are named `--`: | variant | LiDAR | LiDAR noise | IMU noise | shutter | |---|---|---|---|---| | `os1_32-clean-global` | OS1-32 style | none | none | global | | `os1_32-clean-rolling` | OS1-32 style | none | none | rolling | | `os1_32-noisy-global` | OS1-32 style | 2 cm | MEMS | global | | `os1_32-noisy-rolling` | OS1-32 style | 2 cm | MEMS | rolling | | `os1_32-real-global` | OS1-32 style | realistic error model | MEMS | global | | `os1_32-real-rolling` | OS1-32 style | realistic error model | MEMS | rolling | | `vlp16-noisy-rolling` | VLP-16 style | 2 cm | MEMS | rolling | | `os1_64-noisy-rolling` | OS1-64 style | 2 cm | MEMS | rolling | | `os1_64-real-rolling` | OS1-64 style | realistic error model | MEMS | rolling | The `real` variants are the closest to a real sensor and the intended default for benchmarking; the others isolate one factor at a time. The noise seed depends only on the sequence, so the variants of a sequence are paired. The `os1_64` variants differ from their `os1_32` counterparts only in the ring count. | rig | rings | vertical FOV | columns | rate | range | |---|---|---|---|---|---| | `os1_32` | 32, uniform | 45 deg (+-22.5) | 1024 | 10 Hz | 0.5 to 60 m | | `vlp16` | 16, uniform | 30 deg (+-15) | 1808 | 10 Hz | 0.4 to 100 m | | `os1_64` | 64, uniform | 45 deg (+-22.5) | 1024 | 10 Hz | 0.5 to 60 m | - **Realistic error model** (`real` variants, per ray): no return beyond a maximum incidence angle; Gaussian range noise with std `sqrt(s0^2 + (k r)^2)`; the beam footprint spread over oblique surfaces; mixed returns at depth edges; per-ring elevation, azimuth and range biases. The parameters are drawn per sequence and stored next to each rawlog as `.errors.json`. - **Global shutter**: the whole sweep is cast from the pose at the sweep start; every point has `t = 0`. - **Rolling shutter**: column `c` fires at `t = c * 0.1 / columns` s after the sweep start, from the pose at that instant. Points are given in the LiDAR frame at their own firing time (motion-skewed, as real drivers output them); the per-point `t` field lets a consumer de-skew them. - **IMU** (`imu1`): 200 Hz angular velocity and specific force (+9.81 m/s^2 on +Z at rest) in the IMU frame, no orientation output. Noise in the noisy and real variants: | | white noise (per sample) | density | bias random walk | |---|---|---|---| | gyroscope | 3.7e-3 rad/s | 2.6e-4 rad/s/sqrt(Hz) | 2.0e-5 rad/s/sqrt(s) | | accelerometer | 3.2e-2 m/s^2 | 2.3e-3 m/s^2/sqrt(Hz) | 1.0e-3 m/s^2/sqrt(s) | - **Wheel odometry** (`odom`): planar odometry at 10 Hz derived from the ground truth (1% relative noise in the noisy variants). It is written for every carrier, including the hand-held, legged and dog ones, where it is not physically meaningful: leave it out for those, and whenever a LiDAR-only or LiDAR-inertial method is evaluated. - **Intensity**: not generated. Frames: all rotations between sensor frames are identity; `lidar1` is at (0, 0, 0.75) m and `imu1` at (0.05, -0.03, 0.45) m in `base_link`. The ground truth is the pose of `base_link`. Conventions: world frame Z up, gravity 9.81 m/s^2 along -Z; meters, seconds, radians. Timestamps start at `t = 100 s` in every sequence. LiDAR observations are stamped at the sweep start; the per-point `t` is relative to that stamp. TUM files: `timestamp x y z qx qy qz qw`. ## Semantic labels Each point carries two `uint16` fields: `label`, the class of the surface its ray hit, and `instance` (0 for extended surfaces such as roads, terrain, floors and walls). Mixed returns are labeled `outlier`. Outdoor classes reuse the SemanticKITTI raw label ids; indoor ones start at 100. | id | class | id | class | id | class | |---|---|---|---|---|---| | 0 | unlabeled | 50 | building | 100 | floor | | 1 | outlier | 51 | fence | 101 | ceiling | | 10 | car | 52 | other-structure | 102 | wall | | 20 | other-vehicle | 70 | vegetation | 103 | column | | 40 | road | 71 | trunk | 104 | stairs | | 44 | parking | 72 | terrain | 105 | table | | 48 | sidewalk | 80 | pole | 106 | chair | | 49 | other-ground | 81 | traffic-sign | 107 | shelf | | | | 99 | other-object | 108 | other-furniture | | | | | | 109 | goods | ## Files Per sequence, under `//`: | file | contents | |---|---| | `.rawlog` | MRPT rawlog (gzip): `CObservationPointCloud` (`lidar1`, a `CGenericPointsMap` with fields `x y z t ring label instance`; `sensorPose` = `lidar1` in `base_link`), `CObservationIMU` (`imu1`; `sensorPose` = `imu1` in `base_link`), `CObservationOdometry` (`odom`), in time order | | `.gt.tum` | `base_link` ground truth at the sweep start times (10 Hz) | | `.lidar1.gt.tum` | `lidar1` ground truth at the sweep start times | | `.counts.json` | number of sweeps, points and IMU / odometry samples | | `.gen.log` | generator command line and log | | `.errors.json` | `real` variants only: the drawn parameters of the LiDAR error model | | `gt_200hz.tum` | `base_link` ground truth at 200 Hz (the same for every variant) | Some odometry systems stamp a scan at its mid-sweep time: compare against the ground truth at the estimate's own timestamps (e.g. interpolating `gt_200hz.tum`), not by nearest-stamp association with the 10 Hz files. Rawlogs can be read with [MRPT](https://github.com/MRPT/mrpt) (C++ or Python) and run directly with [MOLA](https://github.com/MOLAorg/mola) (`mola-lo-cli-rawlog`). ## Using the data from ROS 2 `rawlog2rosbag` (in [mrpt_ros_bridge](https://github.com/MRPT/mrpt_ros_bridge), branch `ros2`) converts a sequence into a ROS 2 bag: ```bash S=office/spins rawlog2rosbag -i $S/os1_32-real-rolling.rawlog -o office_spins_os1_32-real-rolling \ --gt-tum $S/gt_200hz.tum ``` | topic | type | contents | |---|---|---| | `/lidar1/points` | `sensor_msgs/PointCloud2` | one sweep per message (10 Hz), in the `lidar1` frame: `x y z t` (float32), `ring`, `label`, `instance` (uint16); `t` in seconds from the header stamp | | `/imu1/imu` | `sensor_msgs/Imu` | 200 Hz angular velocity and specific force, in the `imu1` frame (no orientation) | | `/odom/odom` | `nav_msgs/Odometry` | wheel odometry, `odom` to `base_link` (10 Hz) | | `/tf`, `/tf_static` | `tf2_msgs/TFMessage` | `odom` to `base_link`; `base_link` to `lidar1` and `imu1` | | `/ground_truth` | `nav_msgs/Odometry` | with `--gt-tum`: `base_link` in the `map` frame | Stamps start at 100 s: play the bag with `ros2 bag play --clock` and run the nodes with `use_sim_time:=true`. ## Limitations - Geometry is made of what MVSim ray-traces exactly: boxes, walls, cylinders, spheres, wedges, planes and heightfields. Trees are trunks, branches and clusters of spheres; cars are boxes and wheels. - No glass, intensity, multi-echo or reflections. The error model is geometric (no material model). - The scenes are static. - The IMU has white noise and bias random walk, but no initial bias, scale factor or axis misalignment. All sensors share one clock. ## License [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/): you may share and adapt the data for any purpose, including commercial use and model training, with attribution. ## Citation A paper describing the benchmark is in preparation; its reference will be added here. Until then, please cite this dataset page.