README.md

---
pretty_name: pi-embodied Policy Multimodal Interactions
language:
- en
- zh
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
dataset_info:
  features:
  - name: id
    dtype: string
  - name: schema_version
    dtype: string
  - name: trajectory_unit
    dtype: string
  - name: source
    dtype: string
  - name: session_id
    dtype: string
  - name: turn_id
    dtype: string
  - name: model
    dtype: string
  - name: messages
    list:
    - name: role
      dtype: string
    - name: content
      list:
      - name: type
        dtype: string
      - name: text
        dtype: string
      - name: path
        dtype: string
      - name: mime_type
        dtype: string
    - name: tool_calls
      list:
      - name: id
        dtype: string
      - name: type
        dtype: string
      - name: name
        dtype: string
      - name: arguments
        dtype: string
    - name: tool_call_id
      dtype: string
    - name: name
      dtype: string
    - name: timestamp
      dtype: string
  - name: messages_json
    dtype: string
  - name: tools_json
    dtype: string
  - name: tools_status
    dtype: string
  - name: images
    list: image
  - name: media_refs
    sequence: string
  - name: media_status
    dtype: string
  - name: outcome
    dtype: string
  - name: repository_url
    dtype: string
  - name: base_commit
    dtype: string
  - name: patch
    dtype: string
  - name: test_result_json
    dtype: string
  - name: provenance_json
    dtype: string
  - name: native_records_json
    dtype: string
  - name: lerobot_status
    dtype: string
---

# pi-embodied Policy Multimodal Interactions

This public dataset contains 904 project-scoped agent turns, including 848 turns with recoverable, published images. It preserves successful, unsuccessful, interrupted and unverified interactions without inventing outcome labels.

## Storage

- `data/train-*.parquet`: normalized interactions with an explicit Arrow schema and Zstandard compression.
- `media/<hash-prefix>/<sha256>.<extension>`: deduplicated original image bytes, referenced by messages and exposed through the Hugging Face `images` feature.
- `manifest.json`: row counts, media recovery status and image checksums.
- `data/batches/` and `manifests/`: immutable historical text archives, excluded from the default dataset configuration.
- `load_multimodal.py`: loading helpers for interactions, chat messages and execution exports.

## Conversation schema

Each row corresponds to a previously selected task turn, not a certified full robot episode. `messages` stores ordered user/assistant/tool messages, content parts, call IDs, function names, arguments and timestamps. `messages_json` stores an OpenAI-style conversation with `tool_calls.function` and `image_url` content parts. Decode it with `json.loads`; image URLs resolve to this repository. Visible reasoning summaries are retained when available; encrypted provider content is excluded.

`images` is a list of Hugging Face Image features; `media_refs` identifies the corresponding repository-relative files. Original source records, after media extraction and text redaction, are available in `native_records_json`. `provenance_json` links every normalized row to its immutable archive identity. The same row IDs and source classification are retained.

`tools_json` is `[]` and `tools_status` is `not_recorded` where the original collection did not preserve tool definitions. Tool definitions are not reconstructed from argument examples. Repository revisions, patches, test results and task outcomes remain null or unknown when not independently linked to the recorded turn.

## Media recovery and handling

Original inline images were recovered from source logs for previously selected turns. Recoverable explicit local image attachments are included. Referenced files that no longer exist remain marked as missing. Remote URLs are not crawled to reconstruct unavailable content. Text is filtered for recognizable credentials and personal identifiers. Visual outputs of known robot execution and observation tools are retained with their source provenance; other images undergo OCR filtering for recognizable sensitive content. OCR is a filter rather than a guarantee. Original source logs and withheld images remain on their private source/staging storage.

The default `train` split denotes a usable storage split, not a claim that all examples are suitable for supervised training. Outcome-based curation, overlapping-history deduplication, context completion and tool-schema capture remain necessary for individual training recipes. No new redistribution or training license is assigned; original third-party material retains its applicable rights.

## LeRobot v3 execution exports

`lerobot/<export-name>/` contains separate native LeRobot v3 roots, each with `meta/`, `data/` and, when applicable, `videos/`. Different action spaces and coordinate conventions remain separate. These are integration-verification exports, not certified successful demonstrations; the two XPolicy exports represent different action/state views of the same source episode.

`lerobot/inventory.json` records source paths, frame counts, rates and provenance. `lerobot/validation-report.json` records structural checks. `lerobot/agent-links.parquet` records exact source-path mentions in selected agent turns; a path mention is not a verified call-to-frame mapping. No frame association is fabricated where the original recording lacked a linking identifier.

Use `load_execution_export(name)` from the included helper, or download the selected subtree and pass its local root to `LeRobotDataset`. The repository root is the interactions collection and is not itself a single LeRobot dataset. Continuous state/action recordings are never synthesized from sparse agent tool calls.