Datasetpublic

justintiensmith/VLA_Reasoning_Training_Dataset_1200

Spa-Bench fine-tuning demonstrations — full five-camera release This is the canonical full-length demonstration dataset used by Spa-Bench, a real-robot benchmark for spatially grounded reasoning in vision-language-action policies. Explore episodes in the LeRobot visualizer Dataset summary Field Value Episodes 1,200 Frames 612,733 Duration at 30 FPS approximately 5.7 hours Unique instruction strings 321 Task families 6; 200 demonstrations per… See the full description on the dataset page: https://huggingface.co/datasets/justintiensmith/VLA_Reasoning_Training_Dataset_1200.

sourceHugging Faceapache-2.0updated 16d agoView on Hugging Face

0likes202downloads

Dataset Card

Spa-Bench fine-tuning demonstrations — full five-camera release

This is the canonical full-length demonstration dataset used by Spa-Bench, a real-robot benchmark for spatially grounded reasoning in vision-language-action policies.

Explore episodes in the LeRobot visualizer

Dataset summary

Field	Value
Episodes	1,200
Frames	612,733
Duration at 30 FPS	approximately 5.7 hours
Unique instruction strings	321
Task families	6; 200 demonstrations per family
Objects	25 objects across 11 morphological families
Camera streams	middle, wrist, above, left, right
Robot state/action	six absolute SO-101 joint positions
Home-start / recovery demonstrations	883 / 317

The six task families are State Recognition, Relational Placement, Referential Disambiguation, Ordinal Reference Ordering, Counting, and Size Recognition. The missing-cell design withholds selected concept–argument combinations while retaining exposure to their constituent concepts, objects, and broad manipulation behaviours.

Assembly and provenance

—Episodes 0–199 come from the corrected State Recognition recollection set.
—Episodes 200–1199 preserve the physical recordings and order from `justintiensmith/VLA_Benchmark_Prompted_1200`.
—Final instructions are assigned by the benchmark prompt manifest.
—Episode-level design and provenance are stored in meta/training_manifest.parquet and meta/training_manifest.csv.
—Prompt manifests, validation reports, and object metadata are stored under meta/benchmark/.

The immutable data revision used by the pi0.5 and MolmoAct2 runs is `3c92bdd41891803603190b441144274544dab20b`.

Dataset structure

This is a LeRobot v3 dataset at 30 FPS. Each frame contains an episode index, frame index, timestamp, task index, six-dimensional observation state, six-dimensional action, and synchronized 480×640 RGB observations. The evaluated policies use only the middle and wrist views even though all five are retained here.

Uses and limitations

This release is intended for embodied-AI research, benchmark inspection, and reproduction of the Spa-Bench adaptations. It is not representative of general homes or unconstrained manipulation: all data use one SO-101 embodiment, one workspace, one collection setup, a limited object set, and English instructions. The dataset encodes a deliberate compositional holdout and should not be treated as an IID sample of natural robot tasks.

Robot data and learned policies can enable unsafe motion. Use hardware safeguards and human supervision. Review incidental laboratory imagery before redistributing extracted frames.

Related artifacts

Citation

Please cite the completed Spa-Bench MSc report and the thesis artifact.