justintiensmith/VLA_Reasoning_Training_Dataset_1200_2cam

Spa-Bench fine-tuning demonstrations — full two-camera derivative

This is the two-camera projection of the canonical full-length Spa-Bench demonstration dataset. It retains the middle and wrist views consumed by the evaluated policies and omits the unused above, left, and right streams.

Explore episodes in the LeRobot visualizer

Dataset summary

Field	Value
Episodes	1,200
Frames	612,733
Unique instruction strings	321
Frame rate	30 FPS
Camera streams	fixed middle and wrist
Robot state/action	six absolute SO-101 joint positions

Episode identities, frame counts, task instructions, state, and action are preserved from the five-camera dataset. The retained middle and wrist video objects are byte-identical to the canonical release. Metadata is rewritten to remove references to the omitted streams.

The immutable revision used by the GR00T-N1.7 Full Fine-Tune run is b82cdc8f131f33320f38d7a990337c0fd98f1353.

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 from the middle and wrist cameras.

Uses and limitations

Use this derivative when a training pipeline needs only the two evaluated camera views. Use the canonical five-camera release for complete source observations and benchmark provenance.

The dataset comes from one robot embodiment, workspace, collection setup, limited object set, and English instruction distribution. It implements a deliberate compositional holdout and is not an IID sample of natural robot tasks. Robot data and learned policies can enable unsafe motion; use hardware safeguards and human supervision.

Citation

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