What a Human Task Episode Must Preserve
A field checklist for capturing human task episodes that retain the synchronized physical evidence needed for Physical AI training and evaluation.
A task episode is an evidence unit
A human demonstration becomes useful to Physical AI only when the task can be reconstructed. A video file may show a person completing an operation, but it does not guarantee that the goal, state, contact, failure, or recovery can be recovered later.
A human task episode should be treated as an evidence unit with a shared clock, spatial calibration, task context, and a verifiable outcome.
Preserve the task definition
Before capture begins, record the task goal, objects, tools, workspace, constraints, expected result, and success criteria. Without this context, two visually similar actions may represent different tasks.
The task definition should also identify the operator, capture setup, environment revision, and any robot embodiment or downstream model requirements.
Preserve synchronized perception
First-person and external views should be timestamped against the same task clock. Camera intrinsics, extrinsics, lens identity, frame rate, dropped frames, and calibration revisions belong with the episode.
The purpose is not merely replay. Synchronization allows an engineer or model to align what was visible with the action, contact, and state transition that followed.
Preserve motion and action
Capture head pose, hand pose, tool pose, action trajectories, and robot state where available. Motion evidence should retain continuity and coordinate frames rather than only sparse annotations.
For human data, the action representation may later require abstraction or embodiment mapping. Preserving raw geometry keeps those options open.
Preserve force and contact
Contact is where a task begins to change the physical world. Useful evidence may include force/torque, tactile signals, contact events, gripper state, object movement, or state estimates derived from synchronized sensors.
When direct force sensing is unavailable, the episode should still preserve the observations and metadata needed to estimate contact with an explicit confidence level.
Preserve failure and recovery
Do not stop recording immediately when execution diverges. The deviation, operator diagnosis, corrective action, and recovered state often contain more reusable information than a clean success.
Failure evidence should distinguish the observed deviation from the inferred cause. This separation allows the Physical Causal Model to validate explanations against the episode.
Preserve quality and lineage
Each episode should carry calibration status, clock skew, missing streams, quality checks, consent or authorization, version, and provenance. A training asset is only reusable when another team can determine where it came from and whether it satisfies the intended quality threshold.
The Physical AI DBMS organizes these task-native records and preserves versioned lineage.
A minimum episode manifest
A practical manifest should include:
- task and environment identifiers;
- operator and capture-system identifiers;
- start, end, and shared clock information;
- camera and sensor calibration;
- available signal streams and sampling rates;
- task outcome and known failure/recovery events;
- quality results, missing data, and approvals;
- file locations, hashes, versions, and usage rights.
The Task Interface is the production unit that creates this evidence package. The Ego Data pillar page explains the source evidence, while Data Compilation explains how it becomes operational data.
Discuss one concrete evidence package with ArcheBase by requesting a data demo.