# Egocentric Data Collection Brief Version: `[0.1]` Owner: `[name / team]` Last updated: `[YYYY-MM-DD]` Status: `[draft / pilot-approved / scale-approved]` Use this worksheet to define an egocentric dataset before requesting price, volume, or a collection launch. Replace every bracketed field. Mark unknown requirements explicitly; do not silently treat them as flexible. ## 1. Learning or Evaluation Objective - Target model or system: `[ ]` - Training stage or evaluation workflow: `[ ]` - Decision this dataset must support: `[ ]` - Model-facing signal required from the data: `[ ]` - Current baseline or data source: `[ ]` - Measured failure or coverage gap: `[ ]` - Signals this dataset will not provide: `[ ]` - Other data sources that supply the missing signals: `[ ]` ### Success definition Describe how downstream value will be evaluated. Prefer target-model or evaluation metrics over raw collection volume. `[ ]` ## 2. Task and Episode Definition - Task family: `[ ]` - Episode unit: `[one attempt / one procedure / other]` - Start condition: `[ ]` - Successful terminal condition: `[ ]` - Failed terminal condition: `[ ]` - Interruption and retry policy: `[ ]` - Required task steps: `[ ]` - Allowed natural variation: `[ ]` - Required mistakes or recovery behavior: `[ ]` - Unsafe or prohibited behavior: `[ ]` ### Critical visibility moments List every action, object state, contact, tool use, or scene event that must be visible for an episode to be usable. 1. `[ ]` 2. `[ ]` 3. `[ ]` ## 3. Coverage Matrix Define target coverage and how it will be measured. | Dimension | Required values or slices | Target | Pilot minimum | Exclusions | |---|---|---:|---:|---| | Tasks | `[ ]` | `[ ]` | `[ ]` | `[ ]` | | Environments | `[ ]` | `[ ]` | `[ ]` | `[ ]` | | Objects and tools | `[ ]` | `[ ]` | `[ ]` | `[ ]` | | Layouts and lighting | `[ ]` | `[ ]` | `[ ]` | `[ ]` | | Contributor criteria | `[ ]` | `[ ]` | `[ ]` | `[ ]` | | Geography or language | `[ ]` | `[ ]` | `[ ]` | `[ ]` | | Failure and recovery | `[ ]` | `[ ]` | `[ ]` | `[ ]` | Natural frequency versus deliberate balance policy: `[ ]` ## 4. Viewpoint and Capture Hardware - Primary viewpoint: `[head / glasses / chest / wrist / robot / other]` - Reason this viewpoint exposes the required signal: `[ ]` - Additional ego or exo cameras: `[ ]` - Unique information each additional camera contributes: `[ ]` - Device and lens: `[ ]` - Resolution: `[ ]` - Frame rate: `[ ]` - Orientation and field of view: `[ ]` - Stabilization policy: `[ ]` - Audio policy: `[off / on / conditional]` - Fit, comfort, safety, and slippage checks: `[ ]` ### Mount acceptance - Minimum percentage of critical actions visible: `[ ]` - Maximum tolerated occlusion duration: `[ ]` - Framing or horizon tolerance: `[ ]` - Motion and exposure rejection rule: `[ ]` - Population and workspace conditions used in the pilot: `[ ]` ## 5. Modalities and Synchronization For each required modality, define device, units, sampling, timebase, calibration, and missing-data policy. | Modality | Required | Sampling / format | Timebase | Calibration | Missing-data policy | |---|---|---|---|---|---| | RGB video | `[ ]` | `[ ]` | `[ ]` | `[ ]` | `[ ]` | | Audio | `[ ]` | `[ ]` | `[ ]` | `[ ]` | `[ ]` | | Depth | `[ ]` | `[ ]` | `[ ]` | `[ ]` | `[ ]` | | IMU | `[ ]` | `[ ]` | `[ ]` | `[ ]` | `[ ]` | | Gaze | `[ ]` | `[ ]` | `[ ]` | `[ ]` | `[ ]` | | Hand / body pose | `[ ]` | `[ ]` | `[ ]` | `[ ]` | `[ ]` | | Robot state / action | `[ ]` | `[ ]` | `[ ]` | `[ ]` | `[ ]` | | Exocentric cameras | `[ ]` | `[ ]` | `[ ]` | `[ ]` | `[ ]` | - Shared clock or synchronization method: `[ ]` - Maximum tolerated drift: `[ ]` - Dropped-frame policy: `[ ]` - Long-take alignment test: `[ ]` ## 6. Contributor and Field Protocol - Eligibility criteria: `[ ]` - Required experience or task familiarity: `[ ]` - Compensation framework: `[ ]` - Instruction and protocol version: `[ ]` - Training or calibration procedure: `[ ]` - Setup checks before each take: `[ ]` - Repeat conditions: `[ ]` - Immediate safety stop: `[ ]` - Immediate privacy stop: `[ ]` - Device or synchronization stop: `[ ]` - Batch-level quality pause threshold: `[ ]` - Escalation owner: `[ ]` ## 7. Metadata and Episode Schema ### Capture facts - Asset and episode IDs: `[ ]` - Device and configuration: `[ ]` - Timestamps and timebase: `[ ]` - Codec, dimensions, frame rate, and duration: `[ ]` - Checksums and file relationships: `[ ]` ### Context - Task and step identifiers: `[ ]` - Environment and scene attributes: `[ ]` - Viewpoint and mount: `[ ]` - Protocol and instruction version: `[ ]` - Contributor attributes permitted in the dataset: `[ ]` ### Governance - Rights and consent reference: `[ ]` - Privacy review status: `[ ]` - Distribution tier: `[ ]` - Retention and deletion rule: `[ ]` - Release and lineage fields: `[ ]` ### Data contract Attach a schema defining every field name, type, unit, allowed null, vocabulary, relationship, time convention, coordinate frame, and compatibility version. Schema location: `[ ]` ## 8. Annotation Plan - Annotation consumer and model objective: `[ ]` - Unit: `[clip / segment / event / frame / object / other]` - Required labels: `[ ]` - Ontology version: `[ ]` - Guideline version: `[ ]` - Confidence or abstention behavior: `[ ]` - Overlap and disagreement method: `[ ]` - Adjudication owner: `[ ]` - Model-assisted annotation lineage: `[ ]` - Rework and correction policy: `[ ]` ### Annotation acceptance - Quality metric by label type: `[ ]` - Threshold: `[ ]` - Calibration-set composition: `[ ]` - Rare and ambiguous slices: `[ ]` - Sampling and audit frequency: `[ ]` ## 9. Automated Quality Checks Define pass, quarantine, repair, and reject behavior for each check. | Check | Threshold | Failure action | Evidence retained | |---|---|---|---| | File decode and integrity | `[ ]` | `[ ]` | `[ ]` | | Dimensions and frame rate | `[ ]` | `[ ]` | `[ ]` | | Duration and episode bounds | `[ ]` | `[ ]` | `[ ]` | | Timestamps and monotonicity | `[ ]` | `[ ]` | `[ ]` | | Required metadata | `[ ]` | `[ ]` | `[ ]` | | Duplicates | `[ ]` | `[ ]` | `[ ]` | | Modality alignment | `[ ]` | `[ ]` | `[ ]` | | Schema and relationships | `[ ]` | `[ ]` | `[ ]` | ## 10. Human Review and Release Gates - Task completion review: `[ ]` - Critical visibility review: `[ ]` - Motion, framing, and lighting review: `[ ]` - Privacy review: `[ ]` - Annotation fitness review: `[ ]` - Rights and license review: `[ ]` - Dataset-card approval: `[ ]` - Final release owner: `[ ]` Review sampling must cover tasks, environments, devices, protocol versions, contributor cohorts, and time—not only the easiest completed rows. ## 11. Privacy, Rights, and Security - Jurisdictions: `[ ]` - Contributor notice, consent, contract, or other lawful framework: `[ ]` - Bystander policy: `[ ]` - Minors and sensitive-environment policy: `[ ]` - Screens, documents, voices, reflections, and location policy: `[ ]` - Environment preparation and no-record zones: `[ ]` - Redaction and transformation rules: `[ ]` - Raw, annotation, buyer, and public access tiers: `[ ]` - Security and access logging: `[ ]` - Retention and deletion: `[ ]` - Incident and complaint response: `[ ]` - Withdrawal handling where applicable: `[ ]` ### Required commercial rights - Model training: `[ ]` - Evaluation: `[ ]` - Derived models: `[ ]` - Third-party annotation or engineering access: `[ ]` - Retention: `[ ]` - Geographic use: `[ ]` - Redistribution: `[ ]` - Exclusivity: `[ ]` - Public examples or marketing: `[ ]` ## 12. Delivery and Release - Target format and exact version: `[native / RLDS / LeRobot / custom]` - Storage and directory layout: `[ ]` - Media encoding: `[ ]` - Manifest format: `[ ]` - Dataset card: `[ ]` - Splits and leakage controls: `[ ]` - Checksums: `[ ]` - Transformation history: `[ ]` - Known limitations: `[ ]` - License and rights packet: `[ ]` - Release naming and versioning: `[ ]` - Correction and migration policy: `[ ]` ## 13. Buyer Ingest Test - Target loader and version: `[ ]` - Execution environment: `[ ]` - Pilot sample size: `[ ]` - Modalities decoded: `[ ]` - Episode traversal tested: `[ ]` - Batch construction tested: `[ ]` - Representative records manually inspected: `[ ]` - Training or evaluation smoke test: `[ ]` - Failures and resolutions: `[ ]` - Ingest owner and approval date: `[ ]` ## 14. Pilot Design and Scale Decision - Hardest tasks included: `[ ]` - Hardest environments included: `[ ]` - Viewpoint risks exercised: `[ ]` - Privacy conditions exercised: `[ ]` - Annotation risks exercised: `[ ]` - Delivery and loader path exercised: `[ ]` - Pilot acceptance thresholds: `[ ]` - Required evidence packet: `[ ]` ### Decision record - Result: `[approved / revise and repeat / stopped]` - Passed requirements: `[ ]` - Failed requirements: `[ ]` - Unverified requirements: `[ ]` - Protocol changes: `[ ]` - Specification changes: `[ ]` - Remaining risks: `[ ]` - Approved scale and conditions: `[ ]` - Decision owner and date: `[ ]` --- This template is operational guidance, not legal advice. Adapt privacy, consent, labor, security, and licensing requirements with qualified counsel for the actual jurisdictions and program.