Synchronized experience
Observations, motion, actions, task boundaries, and outcomes aligned on a shared timeline.
ZL Corpus
A reusable data product for Physical AI.
Package synchronized observations, motion, actions, task context, provenance, and reviewed labels into versioned corpora built for model training and evaluation.
Data product
The delivery layer of the Zerolaw stack.
ZL Corpus turns processed and reviewed Physical AI episodes into a consistent product that model teams can reuse.
Each release keeps the data, schema, calibration context, provenance, task semantics, and quality state connected. Teams receive a corpus that can be filtered, versioned, split, and traced instead of a directory of unrelated recordings.
Inside a corpus release
The context required to use the data.
A ZL Corpus release carries the relationships that would otherwise have to be reconstructed downstream.
Consistent interfaces
Explicit signal definitions, coordinate systems, timestamps, and dataset structure.
Traceable origin
Capture device, calibration, environment, processing version, and review history.
Reviewed status
Validation results, confidence, known exceptions, and human-reviewed task semantics.
Product architecture
Built through the complete Zerolaw pipeline.
- 01ZL Capture
Records task-relevant vision, inertial, depth, pose, and custom multimodal signals.
- 02ZL Core
Cleans, slices, synchronizes, calibrates, and fuses captured streams.
- 03ZL Annotation
Adds reviewed task semantics and feeds quality findings back into platform rules.
- 04ZL Corpus
Organizes accepted episodes, schemas, metadata, and splits as a versioned data product.
Delivery contract
A corpus release is explicit about what it contains.
| Layer | Example contents | Why it matters |
|---|---|---|
| Episode data | Vision, IMU, pose, depth, actions, task states | Preserves the observation-action sequence |
| Schema | Signal definitions, timestamps, coordinate frames | Keeps training interfaces consistent |
| Metadata | Environment, task, device, subject, outcome | Supports filtering and coverage analysis |
| Quality | Validation status, confidence, exceptions, review | Supports acceptance and sampling decisions |
| Versioning | Release identifier, processing history, split definition | Makes experiments reproducible |
Model development
Designed around repeatable training decisions.
Corpus exploration
Filter episodes by task, environment, signal availability, outcome, and quality state.
Training splits
Keep explicit train, validation, and evaluation definitions tied to a corpus version.
Failure analysis
Trace model behavior back to capture context, calibration, labels, and processing history.
Dataset evolution
Add coverage and corrections without losing the provenance of earlier experiments.
ZL Corpus FAQ
Product questions.
What is ZL Corpus?
ZL Corpus is Zerolaw's Physical AI data product. It packages synchronized episodes, schemas, provenance, reviewed labels, quality information, and dataset splits into a reusable release.
How is it different from raw recordings?
Raw recordings are sensor files. ZL Corpus preserves the relationships between observations, motion, actions, task context, calibration, labels, and outcomes in a consistent versioned structure.
How does it connect to the platform?
ZL Capture records the required signals, ZL Core processes and synchronizes them, ZL Annotation adds reviewed semantics and quality feedback, and ZL Corpus organizes the resulting episodes for model development.
Build with Zerolaw