Observation
RGB, stereo, depth, and other sensor views of the task.
Physical AI data
Real-world experience, structured for training.
Connect observations, motion, actions, geometry, task context, and outcomes in synchronized episodes that models can learn from and teams can evaluate.
Definition
What is Physical AI data?
Physical AI data is structured real-world experience for systems that perceive, reason, and act in physical environments.
Its value comes from the connection between what was observed, what action occurred, how the environment changed, and whether the task succeeded. That connection must survive capture, processing, annotation, and delivery.
The episode contract
Six layers make experience trainable.
Each data program selects the layers required by its model and evaluation target.
Motion
Inertial streams and pose that describe how the actor moves.
Action
Human or robot behavior aligned to the observations it changes.
Geometry
Coordinate systems, maps, depth, and object relationships.
Task
Boundaries, phases, instructions, conditions, and outcomes.
Provenance
Calibration, device, subject, environment, and quality history.
Zerolaw data stack
A controlled path from field capture to corpus.
- 01ZL Capture
Collect task-relevant multimodal signals with synchronized hardware.
- 02ZL Core
Clean, slice, synchronize, calibrate, and fuse each recording.
- 03ZL Annotation
Add task semantics and feed review results back into quality rules.
- 04ZL Corpus
Package consistent episodes, splits, schemas, and provenance for model teams.
Quality model
Volume is useful only after structure is reliable.
| Quality dimension | What it controls | What can be verified |
|---|---|---|
| Task relevance | Whether each signal supports a learning objective | Protocol-to-schema traceability |
| Temporal alignment | Whether cause and effect stay on one timeline | Timestamp drift and sync tolerance |
| Calibration | Whether sensor outputs share usable coordinate systems | Intrinsic, extrinsic, and session checks |
| Coverage | Whether the corpus contains meaningful task variation | Distribution across people, scenes, objects, and outcomes |
| Provenance | Whether errors can be traced and corrected | Device, operator, environment, and processing history |
Training applications
One data foundation, multiple learning paths.
World models
Learn how physical environments, objects, and people evolve through interaction.
Imitation learning
Connect expert demonstrations to observable actions, states, and results.
Physical reinforcement learning
Ground simulation, reward design, policy learning, and offline evaluation in real tasks.
Action fine-tuning
Adapt general models to domain-specific behaviors and operating conditions.
Physical AI data FAQ
Questions model and data teams ask.
What is Physical AI data?
It is structured real-world experience used to train systems that perceive, reason, and act in physical environments. It links observations to motion, actions, geometry, task context, and outcomes.
How is it different from ordinary video?
Video captures appearance. Physical AI data also preserves synchronized sensor signals, calibration, coordinate systems, actions, task boundaries, and results.
What is a Physical AI episode?
An episode is a bounded, synchronized record of a task or interaction, including the observations, motion, actions, context, and result needed for training or evaluation.
What determines data quality?
Task relevance, temporal alignment, calibration, coverage, consistency, provenance, annotation accuracy, and fit with the target training schema.
Build with Zerolaw