Datasets:

CALM-Dataset
/

CALM-Germany

---
license: cc-by-nc-4.0
gated: true
extra_gated_prompt: |
  Please read the following carefully before requesting access.
  By requesting access to this dataset, you confirm that you have read and agree to the dataset license (CC BY-NC 4.0) and the additional usage terms below. Access is provided only under these conditions.

  You agree to:
  - Use the dataset in compliance with the CC BY-NC 4.0 license (non-commercial use only)
  - Use the dataset for research, development, and safety-related purposes
  - NOT use the dataset for military, surveillance, or related applications
  - NOT use the dataset in ways that could cause harm to individuals or society
  - Respect all applicable laws and regulations
  - Cite both CARD and CALM (see Citation below) in any resulting work

extra_gated_fields:
  Full Name: text
  Affiliation (University / Company): text
  Country: text
  Reason for access: text
  I confirm that I have read and will comply with the license and usage terms above: checkbox

task_categories:
  - reinforcement-learning
  - robotics
tags:
  - autonomous-driving
  - ride-comfort
  - speed-planning
  - vehicle-dynamics
  - CARD
pretty_name: CALM-Germany
size_categories:
  - 10K<n<100K
---
# CALM-Germany

CALM-Germany holds 114 speed bump crossings recorded on German roads, cut from the [CARD](https://huggingface.co/CARD-Data) recordings. Each crossing is a short stretch of a drive around a speed bump, with the camera frames, depth points, point cloud and raw sensor files of that stretch, the wheel signal that the road height is built from, and, for the qualified crossings, the speed at which two planners drove over it in simulation. The data is meant for training and evaluating speed planning over speed bumps.

A typical crossing starts 5 s before the bump and ends 5 s after it. Crossings last 4 to 35 s, 10 s in the median: some are cut short where the recording starts or ends, and 15 contain two to five bumps in a row and are longer.

## Crossings

| Area | Crossings | Area | Crossings |
|---|---|---|---|
| Ingolstadt | 20 | Sindelfingen | 14 |
| Munich | 19 | Augsburg | 13 |
| Taufkirchen | 19 | Stuttgart | 12 |
| Ludwigsburg | 16 | Unterföhring | 1 |

The bump in each crossing was measured from the road profile. In the 64 qualified crossings, where this measurement passes the quality check, the bumps are 3.05 to 12.6 cm high and 0.8 to 14.4 m wide. The other 50 are marked `low_confidence` in CALM-Scenarios, because no clear bump was found near the position the camera flagged. The two figures show every crossing with the measured values, before CALM-Scenarios limits them to the range its road model accepts. The first leaves out the 14 crossings wider than 8 m. The second shows the height divided by the width, which the figures call severity.

![Bump height against bump width for each crossing](stats/height_vs_width.svg)

![Bump height divided by width for each crossing, by area](stats/severity_by_area.svg)

`manifest.json` lists every crossing with the CARD sequence it was cut from, its number of bumps and its duration.

## Structure

```
manifest.json                   list of all crossings
stats/                          the figures in this README
sequences/<area>/<crossing>/
├── img/cam_0/, img/cam_1/      camera frames
├── agg_depth/cam_1/            aggregated depth points in the camera frame
├── labels/                     bump box per cam_1 frame
├── raw/
│   ├── gnss.ubx                raw GNSS log
│   └── data.pcap               raw sensor packet capture
├── export/
│   ├── output.laz              point cloud
│   ├── output.laz.index.json   point cloud scan index
│   ├── wheel_excitement.json   distance driven and contact points of each wheel
│   ├── output.laz.trajectory.json      pose of the car
│   └── output.laz.sensor_export.json   GPS and IMU samples
├── comfort_profiles/           only for the 64 qualified crossings
│   ├── ppo.json
│   └── mpc.json
└── meta.yaml                   window and bump position in the source sequence
```

Five crossings, `ludwigsburg_sb04`, `ludwigsburg_sb05`, `ludwigsburg_sb07`, `ludwigsburg_sb08` and `sindelfingen_sb14`, have no `data.pcap`. The LiDAR capture of their source recording ends before the window starts, so no packets exist for it, and their `meta.yaml` lists `pcap: unavailable`.

For seven crossings the label folder holds more files than the image folders, 206 label files in all without a matching camera frame: `ludwigsburg_sb09`, `munich_sb12`, `munich_sb13`, `munich_sb15`, `munich_sb18`, `sindelfingen_sb09` and `sindelfingen_sb13`.

## File formats

**Labels.** CARD's labels exist for `cam_1` only. Each file in `labels/` belongs to the `cam_1` frame with the same timestamp and has one line per box in the YOLO format: class id, then the box centre x and y, width and height as fractions of the image size, for example `1 0.512338 0.504124 0.047686 0.004389`. Class 1 is a speed bump; 3 lines in this dataset have class 0, which CALM does not use. An empty file means that the frame shows no bump.

**Depth.** Each `.npz` file in `agg_depth/cam_1/` holds aggregated depth points for one frame: a single array `pts_cam` of shape N x 3 (float32) with x, y and z of each point in the camera frame. Two files hold no points, `ingolstadt/ingolstadt_sb05/agg_depth/cam_1/cam_1_71002434.npz` and `taufkirchen/taufkirchen_sb19/agg_depth/cam_1/cam_1_81801807.npz`; they are empty in the source copy as well.

**Wheel signal.** CARD's `export/wheel_excitement.json` has one entry per wheel under `wheel_results` (`cariad_wheel_FL`, `_FR`, `_RL` and `_RR`). Its `wheel_excitement` list has one row per centimetre driven, with 4 numbers: the timestamp in microseconds, a second timestamp in microseconds on another clock (a constant offset from the first), the distance driven in metres, and the vertical excitation of the wheel in metres. `contact_point_trajectory.poses` gives, at the same timestamps, the pose of the point where the wheel touches the road as `[timestamp, [x, y, z, qw, qx, qy, qz]]`. CALM builds the road height in CALM-Scenarios from the contact point heights of the two front wheels, placed along the road by the distance column, and replays the recorded speed from the time and distance of the front left wheel.

**meta.yaml.** Timestamps are in microseconds on the clock of the recording, the same clock as in the frame file names. Distances are in metres.

| Key | Meaning |
|---|---|
| `source_sequence` | The CARD recording the crossing was cut from |
| `source_window_index` | Number of the bump encounter within that recording |
| `window_index` | Number of the crossing within its area, as in the folder name |
| `sb_count` | Number of bumps in the crossing |
| `encounter` | The `cam_1` frames in which the labels show the bump: `start_ts`, `end_ts` and `n_frames` |
| `anchor_ts` | Time of the bump, the end of the encounter |
| `window.start_ts`, `window.end_ts` | Start and end of the crossing |
| `clipped_pre`, `clipped_post` | True when the window starts later or ends earlier than planned, for example because the recording starts or ends there |
| `s_bump` | Distance driven from the start of the recording to the bump |
| `s_span` | Distance driven within the window |
| `contents` | Number of items copied for each part: files, poses, rows, scans or samples; `pcap` is `unavailable` when the recording has no packets for the window |

## Speed profiles

`comfort_profiles/` holds two simulated drives over the crossing. `ppo.json` comes from the trained PPO policy of CALM and `mpc.json` from a model predictive controller that plans with the same objective. Each file gives the travel time and a list of samples, each with the time `t_s` in seconds, the distance along the road `s_m` in metres and the speed `v_kmh` in km/h.

## What CALM added

Most of what a crossing holds comes from CARD-Germany, cut to the window: the camera frames, the aggregated depth points, the bump labels, the point cloud, the trajectory of the car, the wheel signal in `export/wheel_excitement.json`, and the raw GNSS log and sensor packets.

CARD recorded long drives. CALM cut each crossing out of them, 5 s before and 5 s after the bump. Bumps that follow each other closely stay together in one longer window, so a crossing can hold up to five bumps. Where a recording starts or ends, the window is shorter.

![How crossings are cut from a recording and what CALM adds to them](stats/crossing_windows.svg)

Besides the cut, CALM added three things:

| Addition | What it is | What it is for |
|---|---|---|
| `meta.yaml` | Where the window lies in the source recording and where the bump is | Finding the crossing again in the CARD sequence |
| `manifest.json` | The list of all crossings with the CARD sequence, number of bumps and duration of each | Finding a crossing without opening every folder |
| `comfort_profiles/` | Simulated speed over the crossing from the PPO policy and the MPC planner, for the qualified crossings only | Showing how a planner slows down before the bump and speeds up again |

## Use

The compact scenarios derived from these crossings, with the road height, the bump and the camera detections, are in [CALM-Scenarios](https://huggingface.co/datasets/CALM-Dataset/CALM-Scenarios), together with the training and evaluation split. The code is in the [CALM repository](https://github.com/CALM-Dataset/CALM). `calm reproduce` trains and evaluates the policy on the scenarios. Set `CALM_SEQUENCES` to the `sequences` folder of this dataset to compare the result with the recorded drives.

## License

CC BY-NC 4.0. Non-commercial use only. The material copied unchanged from CARD-Germany is released by its authors under CC BY 4.0. The additions by CALM (`meta.yaml`, `manifest.json` and the speed profiles) are released under CC BY-NC 4.0.

## Citation

Please cite CARD, which recorded the drives, and CALM.

```bibtex
@inproceedings{elazab2025card,
  title     = {CARD: A Multi-Modal Automotive Dataset for Dense 3D Reconstruction in Challenging Road Topography},
  author    = {Elazab, Gasser and Neuhaus, Frank and Ko{\ss}, Tilman and Splietker, Malte and Date, Aditya and Unterreiner, Michael and Jansen, Maximilian and Hellwich, Olaf},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026},
  eprint    = {2605.05014},
  archivePrefix = {arXiv}
}
```

```bibtex
@misc{calm_dataset,
  title  = {CALM: Comfort-Aware Longitudinal Motion Planning},
  author = {Mohammed El Said Azab Abdelazim},
  year   = {2026},
  url    = {https://huggingface.co/CALM-Dataset}
}
```

	---
	license: cc-by-nc-4.0
	gated: true
	extra_gated_prompt: \|
	Please read the following carefully before requesting access.

	By requesting access to this dataset, you confirm that you have read and agree to the dataset license (CC BY-NC 4.0) and the additional usage terms below. Access is provided only under these conditions.

	You agree to:
	- Use the dataset in compliance with the CC BY-NC 4.0 license (non-commercial use only)
	- Use the dataset for research, development, and safety-related purposes
	- NOT use the dataset for military, surveillance, or related applications
	- NOT use the dataset in ways that could cause harm to individuals or society
	- Respect all applicable laws and regulations
	- Cite both CARD and CALM (see Citation below) in any resulting work

	extra_gated_fields:
	Full Name: text
	Affiliation (University / Company): text
	Country: text
	Reason for access: text
	I confirm that I have read and will comply with the license and usage terms above: checkbox

	task_categories:
	- reinforcement-learning
	- robotics
	tags:
	- autonomous-driving
	- ride-comfort
	- speed-planning
	- vehicle-dynamics
	- CARD
	pretty_name: CALM-Germany
	size_categories:
	- 10K<n<100K
	---

	# CALM-Germany

	CALM-Germany holds 114 speed bump crossings recorded on German roads, cut from the [CARD](https://huggingface.co/CARD-Data) recordings. Each crossing is a short stretch of a drive around a speed bump, with the camera frames, depth points, point cloud and raw sensor files of that stretch, the wheel signal that the road height is built from, and, for the qualified crossings, the speed at which two planners drove over it in simulation. The data is meant for training and evaluating speed planning over speed bumps.

	A typical crossing starts 5 s before the bump and ends 5 s after it. Crossings last 4 to 35 s, 10 s in the median: some are cut short where the recording starts or ends, and 15 contain two to five bumps in a row and are longer.

	## Crossings

	\| Area \| Crossings \| Area \| Crossings \|
	\|---\|---\|---\|---\|
	\| Ingolstadt \| 20 \| Sindelfingen \| 14 \|
	\| Munich \| 19 \| Augsburg \| 13 \|
	\| Taufkirchen \| 19 \| Stuttgart \| 12 \|
	\| Ludwigsburg \| 16 \| Unterföhring \| 1 \|

	The bump in each crossing was measured from the road profile. In the 64 qualified crossings, where this measurement passes the quality check, the bumps are 3.05 to 12.6 cm high and 0.8 to 14.4 m wide. The other 50 are marked `low_confidence` in CALM-Scenarios, because no clear bump was found near the position the camera flagged. The two figures show every crossing with the measured values, before CALM-Scenarios limits them to the range its road model accepts. The first leaves out the 14 crossings wider than 8 m. The second shows the height divided by the width, which the figures call severity.

	![Bump height against bump width for each crossing](stats/height_vs_width.svg)

	![Bump height divided by width for each crossing, by area](stats/severity_by_area.svg)

	`manifest.json` lists every crossing with the CARD sequence it was cut from, its number of bumps and its duration.

	## Structure

	```
	manifest.json list of all crossings
	stats/ the figures in this README
	sequences/<area>/<crossing>/
	├── img/cam_0/, img/cam_1/ camera frames
	├── agg_depth/cam_1/ aggregated depth points in the camera frame
	├── labels/ bump box per cam_1 frame
	├── raw/
	│ ├── gnss.ubx raw GNSS log
	│ └── data.pcap raw sensor packet capture
	├── export/
	│ ├── output.laz point cloud
	│ ├── output.laz.index.json point cloud scan index
	│ ├── wheel_excitement.json distance driven and contact points of each wheel
	│ ├── output.laz.trajectory.json pose of the car
	│ └── output.laz.sensor_export.json GPS and IMU samples
	├── comfort_profiles/ only for the 64 qualified crossings
	│ ├── ppo.json
	│ └── mpc.json
	└── meta.yaml window and bump position in the source sequence
	```

	Five crossings, `ludwigsburg_sb04`, `ludwigsburg_sb05`, `ludwigsburg_sb07`, `ludwigsburg_sb08` and `sindelfingen_sb14`, have no `data.pcap`. The LiDAR capture of their source recording ends before the window starts, so no packets exist for it, and their `meta.yaml` lists `pcap: unavailable`.

	For seven crossings the label folder holds more files than the image folders, 206 label files in all without a matching camera frame: `ludwigsburg_sb09`, `munich_sb12`, `munich_sb13`, `munich_sb15`, `munich_sb18`, `sindelfingen_sb09` and `sindelfingen_sb13`.

	## File formats

	Labels. CARD's labels exist for `cam_1` only. Each file in `labels/` belongs to the `cam_1` frame with the same timestamp and has one line per box in the YOLO format: class id, then the box centre x and y, width and height as fractions of the image size, for example `1 0.512338 0.504124 0.047686 0.004389`. Class 1 is a speed bump; 3 lines in this dataset have class 0, which CALM does not use. An empty file means that the frame shows no bump.

	Depth. Each `.npz` file in `agg_depth/cam_1/` holds aggregated depth points for one frame: a single array `pts_cam` of shape N x 3 (float32) with x, y and z of each point in the camera frame. Two files hold no points, `ingolstadt/ingolstadt_sb05/agg_depth/cam_1/cam_1_71002434.npz` and `taufkirchen/taufkirchen_sb19/agg_depth/cam_1/cam_1_81801807.npz`; they are empty in the source copy as well.

	Wheel signal. CARD's `export/wheel_excitement.json` has one entry per wheel under `wheel_results` (`cariad_wheel_FL`, `_FR`, `_RL` and `_RR`). Its `wheel_excitement` list has one row per centimetre driven, with 4 numbers: the timestamp in microseconds, a second timestamp in microseconds on another clock (a constant offset from the first), the distance driven in metres, and the vertical excitation of the wheel in metres. `contact_point_trajectory.poses` gives, at the same timestamps, the pose of the point where the wheel touches the road as `[timestamp, [x, y, z, qw, qx, qy, qz]]`. CALM builds the road height in CALM-Scenarios from the contact point heights of the two front wheels, placed along the road by the distance column, and replays the recorded speed from the time and distance of the front left wheel.

	meta.yaml. Timestamps are in microseconds on the clock of the recording, the same clock as in the frame file names. Distances are in metres.

	\| Key \| Meaning \|
	\|---\|---\|
	\| `source_sequence` \| The CARD recording the crossing was cut from \|
	\| `source_window_index` \| Number of the bump encounter within that recording \|
	\| `window_index` \| Number of the crossing within its area, as in the folder name \|
	\| `sb_count` \| Number of bumps in the crossing \|
	\| `encounter` \| The `cam_1` frames in which the labels show the bump: `start_ts`, `end_ts` and `n_frames` \|
	\| `anchor_ts` \| Time of the bump, the end of the encounter \|
	\| `window.start_ts`, `window.end_ts` \| Start and end of the crossing \|
	\| `clipped_pre`, `clipped_post` \| True when the window starts later or ends earlier than planned, for example because the recording starts or ends there \|
	\| `s_bump` \| Distance driven from the start of the recording to the bump \|
	\| `s_span` \| Distance driven within the window \|
	\| `contents` \| Number of items copied for each part: files, poses, rows, scans or samples; `pcap` is `unavailable` when the recording has no packets for the window \|

	## Speed profiles

	`comfort_profiles/` holds two simulated drives over the crossing. `ppo.json` comes from the trained PPO policy of CALM and `mpc.json` from a model predictive controller that plans with the same objective. Each file gives the travel time and a list of samples, each with the time `t_s` in seconds, the distance along the road `s_m` in metres and the speed `v_kmh` in km/h.

	## What CALM added

	Most of what a crossing holds comes from CARD-Germany, cut to the window: the camera frames, the aggregated depth points, the bump labels, the point cloud, the trajectory of the car, the wheel signal in `export/wheel_excitement.json`, and the raw GNSS log and sensor packets.

	CARD recorded long drives. CALM cut each crossing out of them, 5 s before and 5 s after the bump. Bumps that follow each other closely stay together in one longer window, so a crossing can hold up to five bumps. Where a recording starts or ends, the window is shorter.

	![How crossings are cut from a recording and what CALM adds to them](stats/crossing_windows.svg)

	Besides the cut, CALM added three things:

	\| Addition \| What it is \| What it is for \|
	\|---\|---\|---\|
	\| `meta.yaml` \| Where the window lies in the source recording and where the bump is \| Finding the crossing again in the CARD sequence \|
	\| `manifest.json` \| The list of all crossings with the CARD sequence, number of bumps and duration of each \| Finding a crossing without opening every folder \|
	\| `comfort_profiles/` \| Simulated speed over the crossing from the PPO policy and the MPC planner, for the qualified crossings only \| Showing how a planner slows down before the bump and speeds up again \|

	## Use

	The compact scenarios derived from these crossings, with the road height, the bump and the camera detections, are in [CALM-Scenarios](https://huggingface.co/datasets/CALM-Dataset/CALM-Scenarios), together with the training and evaluation split. The code is in the [CALM repository](https://github.com/CALM-Dataset/CALM). `calm reproduce` trains and evaluates the policy on the scenarios. Set `CALM_SEQUENCES` to the `sequences` folder of this dataset to compare the result with the recorded drives.

	## License

	CC BY-NC 4.0. Non-commercial use only. The material copied unchanged from CARD-Germany is released by its authors under CC BY 4.0. The additions by CALM (`meta.yaml`, `manifest.json` and the speed profiles) are released under CC BY-NC 4.0.

	## Citation

	Please cite CARD, which recorded the drives, and CALM.

	```bibtex
	@inproceedings{elazab2025card,
	title = {CARD: A Multi-Modal Automotive Dataset for Dense 3D Reconstruction in Challenging Road Topography},
	author = {Elazab, Gasser and Neuhaus, Frank and Ko{\ss}, Tilman and Splietker, Malte and Date, Aditya and Unterreiner, Michael and Jansen, Maximilian and Hellwich, Olaf},
	booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
	year = {2026},
	eprint = {2605.05014},
	archivePrefix = {arXiv}
	}
	```

	```bibtex
	@misc{calm_dataset,
	title = {CALM: Comfort-Aware Longitudinal Motion Planning},
	author = {Mohammed El Said Azab Abdelazim},
	year = {2026},
	url = {https://huggingface.co/CALM-Dataset}
	}
	```