Skip to content

training-data.md

Latest commit

 

History

History
76 lines (60 loc) · 2.88 KB

training-data.md

File metadata and controls

76 lines (60 loc) · 2.88 KB
Raw

Training Data Creation

OEBP training data is created from validated behavior, capability, runtime, trace, episode, and provenance documents. Raw observations, videos, and low-level action streams stay in their original storage system. OEBP records refer to them through stable observation_ref and action_ref values.

This keeps semantic training labels portable while avoiding duplicated media, hidden licensing changes, or accidental mixing of trusted and untrusted data.

Data Creation Flow

  1. Create or collect an embodied episode in a simulator, mock runtime, or robot dataset.
  2. Validate the behavior document, capability profile, invocation result, trace spans, and provenance record.
  3. Build an EpisodeAnnotation that links:
    • behavior ID;
    • capability profile ID;
    • source dataset ID;
    • trace reference;
    • observation reference;
    • action reference;
    • action codec;
    • outcome;
    • quality labels;
    • provenance.
  4. Run scripts/create_training_data.py to emit JSONL training views.
  5. Review generated records before publishing or using them for model training.

Reference Command

python3 scripts/create_training_data.py \
  --manifest datasets/synthetic/v0.1/manifest.json \
  --behavior examples/pick-and-place.behavior.json \
  --capability examples/generic-mobile-manipulator.capability.json \
  --output-dir /tmp/oebp-training

The command creates:

  • planner.jsonl: context and capability input to target behavior graph.
  • next_skill.jsonl: behavior history to next semantic skill and parameters.
  • contract.jsonl: contract condition records for preconditions, invariants, success conditions, and failure conditions.
  • recovery.jsonl: error-code and trace context to recovery behavior.
  • success_estimation.jsonl: observation and contract context to outcome labels.
  • index.json: counts, source manifest, and generated file list.

Safety And Trust Rules

  • Generated training rows MUST NOT embed raw observations, videos, or action streams.
  • Model-generated behaviors are untrusted until schema, semantic, capability, runtime, and provenance gates pass.
  • A successful mock run is evidence for protocol semantics, not physical safety.
  • Dataset releases should preserve source dataset IDs, licenses, and provenance.
  • Human review is required before generated data is promoted into public training datasets.

Training Views

Planner records teach a model to propose a behavior graph for a context, capability profile, and goal.

Next-skill records teach a model to choose the next semantic skill and parameters from behavior history.

Contract records teach a model or evaluator which predicates must hold before, during, or after behavior execution.

Recovery records teach bounded recovery choices for known failure codes.

Success-estimation records teach outcome prediction from observation references, behavior contracts, and quality labels.