vps2a/open_legwheel

Welcome to the main dataset of the Open LegWheel Project!

This project aims to address the lack of traversability data for training future hybrid exploration rovers. It is an open-source project currently maintained by @vps2a (Filip Szczebak). If you have any suggestions or questions, please feel free to reach out at filip@szczebak.com.

This dataset is linked to the Open LegWheel project on GitHub.

What does this dataset contain (in human terms)?

This dataset was built from just over 50 runs at the Kąkolewo Analogue Lunar Terrain campus in Kąkolewo, Poland, managed by Poznań University of Technology.

The runs were carried out on lunar simulant with mixed sandy and rocky terrain, along a circular track laid out around the rover. Earlier runs (L1 configurations) used an omnidirectional overhead light. L2 configurations used lighting that mimicked low-angle sunlight on the lunar surface, with a colour temperature of 5,900 K.

Details of the rig, its rear weight offload, and the components used can be found in the individual experiment configuration files. If any information is missing, please contact me.

The dataset (in machine terms)

This export contains 2275 selected terrain packets. No packets have been removed for overlap. Labels may be null. Training, validation and test membership has not been assigned: the pool is explicitly named unassigned. Choose and document a dataset license before publishing this directory.

Loading

from datasets import load_dataset

packets = load_dataset("parquet", data_files="packets/*.parquet")["train"]
runs = load_dataset("parquet", data_files="runs.parquet")["train"]
# The local generic Parquet loader calls its default pool "train"; that name
# does not establish a scientific train/test split. After uploading this folder:
# packets = load_dataset("your-account/your-dataset", "packets", split="unassigned")

The packet files include Hugging Face feature metadata. RGB, depth and masks are embedded lossless PNG images; no source paths are needed. depth_scale_m converts stored uint16 depth to metres. Mask PNG values are 0 (outside) and 255 (inside). Keep masks distinct from missing/invalid depth pixels.

Packet columns

• export_id is the globally scoped reference used by overlap vectors. packet_id and packet_number retain the original ID and chronological, 1-based source-manifest position. Numbers can have gaps after selection.
• source_id joins to runs.parquet. Experiment/run and processing IDs are also available directly. Only one processing version of each run is allowed.
• class_id uses 0: Firm Terrain; 1: Loose Terrain; 2: Firm to Loose Transition; 3: Loose to Firm Transition. Null means unlabelled, never class zero. classification_method and model_id preserve annotation provenance.
• terrain_and_light_ID is preserved. For KK_01_L1, laboratory_id is KK, terrain_setup_id is KK_01, and lighting_id is L1. Missing IDs remain null. Keep the same physical setup together when assigning evaluation splits, including recordings made under different lighting.
• local_values and spectral_values have shape [time, channel]. Their corresponding boolean *_valid masks and integer *_relative_times_ns are preserved. Relative timestamps are centred on contact_time_ns. All absolute timestamps are integer nanoseconds. Original float64 precision is retained; nonfinite values become null, not zero. channel_names and channel_units define column order. Angular quantities use radians, effort N*m and linear acceleration m/s^2 (per-channel power units are the square of these units).
• spectral_power is the saved PSD [frequency, channel], in channel-units²/Hz. frequencies_hz gives its frequency grid. dominant_frequency_hz and dominant_power have shape [peak rank, channel]; dominant power is a PSD ordinate, not integrated band power.
• band_power has shape [band, channel], in channel-units². Bands have width 10 Hz, with actual edges in band_edges_hz. Integration is the discrete sum of PSD bin values times FFT frequency spacing. DC is excluded. Bins belong to [lower, upper), with the Nyquist endpoint included in the final band. An odd-length FFT need not have a bin exactly at Nyquist. There is no interpolation, zero-padding, spectral recomputation or relative normalization. Invalid/empty bands and bands extending beyond a source's Nyquist frequency are null. Bands share edges across this export, even when source sample rates differ. Inspect per-row coverage before comparing them.
• rgb_patch, depth_patch, their masks and *_crop_origin_uv preserve the selected image patches. geometry contains projected tracks and physical patch boundaries. Full-image calibration is retained in metadata_json.
• Plane slopes, plane normal/offset, fit metrics and IMU uncertainty are accessible as numerical columns. imu_uncertainty_deg is the saved two-sigma pitch uncertainty. metadata_json and features_json preserve additional original fields. Image/depth-derived features are supporting measurements, not inputs to the telemetry-only benchmark unless explicitly declared.
• Full RGB/depth frame columns included: False. Numerical point_map included: False.

Overlap and maximum subset counts

FFT time bounds are contact time plus the first/last saved spectral timestamps. Intervals are closed: shared endpoints count as overlap. References are symmetric, exclude self, and include only other selected packets from the same run. temporal_overlap_ids/spatial_overlap_ids reference export_id; the matching *_overlap_packet_numbers retain the source packet numbers in the same order.

FFT boundary angles are interpolated between corrected encoder ticks using relative integer time differences. Conservative lower/upper unwrapped angle bounds include the surrounding tick positions. They assume monotonic movement between ticks; unrecorded between-tick reversals and calibration errors are not bounded by this method. Detected reversals, out-of-range windows and movement-event intervals greater than 2 seconds block export. Exactly 2 seconds is allowed. This is an exporter-only interpolation limit, saved as maximum_encoder_gap_ns in the run table; it is independent of the pitch estimator's message-gap threshold. Sparse count changes can reflect slow motion and do not establish a recording dropout. Original message-gap status cannot be inferred from these event times. No extrapolation or gap filling is performed.

Spatial overlap compares conservative wheel-travel arcs modulo 2*pi, including repeated laps and arcs crossing zero. Coverage of a whole revolution overlaps every packet. This is the track traversed during the FFT window, not a camera footprint or the full robot footprint. Angular comparisons use a 1e-12 rad tolerance. Spatial comparisons are only within each run, not across runs with potentially different angular origins.

Each run records the exact maximum temporally non-overlapping subset count (earliest-finish interval scheduling) and exact maximum spatially non-overlapping subset count (circular-arc scheduling by trying each anchor). These are separate optimizations, may describe different subsets, and are not estimates of statistically independent sample counts. No subset is filtered or labelled as preferred. Counts describe the exported selection, not all source packets.

Reproducibility

runs.parquet preserves source configurations/provenance in provenance_json, with machine-local source paths and processing creation times omitted. export_manifest.json records settings, class counts, selected references, run summaries and Parquet SHA-256 checksums. The directory hash covers normalized exported Arrow records, settings, run configuration and exporter version. It is independent of selection order, destination, display name and export time. It is not a checksum of Parquet compression bytes. The dependency lockfile and exporter version identify the serializer used to calculate it.