vla-model/construction-syn-data

Construction-Site Multimodal Dataset (CMAD)

Real-world RGB + LiDAR + pose data collected on a live construction site with a custom ground-robot sensor rig. Three sequences, 16,401 synchronized frames (~78 minutes) of camera, LiDAR, and 6-DoF odometry, laid out in a directory/loader convention compatible with the public RELLIS-3D benchmark, so existing off-road traversability-estimation tooling built around that convention can consume it with only a root-path change.

This dataset was collected to support research on self-supervised traversability estimation in tight, unstructured, real-world spaces — construction sites in particular, where narrow corridors, active equipment, and LiDAR-hazardous surfaces (open trenches, standing water, reflective debris) are common. Collection intentionally includes routes an operator judged difficult or borderline-traversable, not only easy/typical driving.

Sensor rig

• RGB camera, 640×480, intrinsics fx≈399.7, fy≈399.8, cx≈330.4, cy≈197.7.
• Solid-state LiDAR, extrinsically calibrated against the rig's static TF tree.

Dataset structure

Sequence	Frames	Duration	Notes
00000	9,767	≈47.0 min
00001	3,100	≈14.6 min
00002	3,534	≈16.5 min	odometry diverges in the final ~3% of frames — see Known Issues below
Total	16,401	≈78.1 min	≈3.5 Hz camera/LiDAR rate

Each sequence directory contains:

<seq>/
├── calib.txt
├── camera_info.txt
├── camera_times.txt
├── frame_times.txt
├── poses.txt                      # per-frame 6-DoF pose, row-major flattened 3x4 [R|t]
├── imu.npy
├── sync_report.csv
├── pylon_camera_node/              # RGB images, one per frame
└── os1_cloud_node_kitti_bin/       # LiDAR point clouds, one per frame (KITTI .bin format)

Known issues

Sequence 00002's onboard LiDAR odometry is reliable for the great majority of the sequence but diverges sharply in its final ~3% of frames (single-frame position jumps of tens of meters, consistent with an odometry tracking failure). Reported here transparently rather than filtered out — if using pose data from this sequence, exclude the tail or verify pose continuity before relying on it.

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

The recordings may be used, shared, and adapted for non-commercial purposes with appropriate attribution. Commercial use is not permitted without separate permission from the dataset authors.