Dataset · Preview Release
DexXperience Preview is open for early access
A multimodal, 4D motion dataset preview built from real DexCapture sessions, distributed with quality scores and lineage.
TL;DR
- DexXperience Preview is the first dataset slice published from our DexCapture program.
- Every sample carries quality scores, lineage records, and consent metadata, not just raw streams.
- Access is by request so we can match dataset slices to model stage and target tasks.
What we mean by preview
Preview is not a brand label. It is a real constraint. The cohort is intentionally narrow: a curated set of tasks, a small operator pool, and a fixed environment list. Teams can train against it today, and we will expand the cohort by month rather than by year.
We focus the preview on quality rather than headline volume. That is a deliberate trade. Models trained on noisy data degrade silently; models trained on aligned, scored, lineage-tracked data fail loudly when they fail, which is much easier to debug.
Signal stack
Each episode is synchronized across vision, motion, spatial sensing, and language metadata. We treat language as a first-class modality so VLA training pipelines can ingest the dataset without re-annotation.
Every record carries the upstream provenance: which DexCapture session it came from, which DexProcess pipeline version produced it, and which quality thresholds were applied. This is how we avoid silent regressions when the pipeline evolves.
How access works
Access is request-based and free for research use. Teams share their model stage, target tasks, and any constraints around environments or operator profiles, and we match a dataset slice to the request. Commercial licensing is available on request and includes broader environment and task coverage.
We deliberately did not put the preview behind a generic download button. The point of the preview is to talk to the teams using it so we know which slices to expand next.
Preview cohort at a glance
- Stage
- Preview cohort, expanding monthly
- Modalities
- RGB · Depth · IMU · Audio · Pose · Language
- Environments
- Logistics · Food prep · Light assembly
- Distribution
- S3 signed URLs and curated samples
- Per-sample metadata
- Quality score · Lineage · Consent flags
- License
- Free for research; commercial on request
Cohort size and environment coverage update with each monthly release.
What this means for you
Request a slice that fits your model
If you are training a world model, VLA stack, or imitation learning policy and need real-world episodes that survive scrutiny on quality and lineage, the preview is open today. Tell us your stage and target tasks and we can scope a slice.
See DexXperience Preview