saaduddinM/OXE_cmu_stretch_embeddings

Language Table (LeRobot) — Embedding-Only Release (DINOv3 + SigLIP2 image features; EmbeddingGemma task-text features)

This repository packages a re-encoded variant of IPEC-COMMUNITY/cmu_stretch_lerobot where raw videos are replaced by fixed-length image embeddings, and task strings are augmented with text embeddings. All indices, splits, and semantics remain consistent with the source dataset while storage and I/O are substantially lighter. To make the dataset practical to upload/download and stream from the Hub, we also consolidated tiny per-episode Parquet files into N large Parquet shards under a single data/ folder. The file meta/sharded_index.json preserves a precise mapping from each original episode (referenced by a normalized identifier of the form data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet) to its shard path and row range, so you keep original addressing without paying the small-file tax.

• Robot: hello_stretch
• Modalities kept: states, actions, timestamps, frame/episode indices, image embeddings, task-text embeddings
• Removed:
• observation.images.image
• License: apache-2.0 (inherits from source)

---

Quick Stats

From meta/info.json and meta/task_text_embeddings_info.json:

• Episodes: 135
• Frames: 25,016
• Tasks (unique): 5
• Chunks (original layout): 1 (chunks_size=1000)
• Shards (this release): 64 Parquet files under data/ (see meta/sharded_index.json)
• FPS: 10
• Image embeddings (per frame):
  • observation.images.image_dinov3 → float32 [1024] (DINOv3 ViT-L/16 CLS)
  • observation.images.image_siglip2 → float32 [768] (SigLIP2-base)
• Task-text embeddings (per unique task):
  • embedding → float32 [768] from google/embeddinggemma-300m
  • Count: 5 rows (one per task)

Note: This is an embedding-only package. The original pixel arrays listed under “Removed” are dropped.

---

Contents

. 
|-- meta/
|   |-- info.json
|   |-- sharded_index.json
|   |-- tasks.jsonl
|   |-- episodes.jsonl
|   `-- task_text_embeddings_info.json
|-- data/
|   |-- shard-00000-of-000NN.parquet
|   |-- shard-00001-of-000NN.parquet
|   |-- ...
|   `-- task_text_embeddings.parquet
`-- README.md

---

How This Was Generated (Reproducible Pipeline)

1. Episode → Image Embeddings (drop pixels) convert_lerobot_to_embeddings_mono.py (GPU-accelerated preprocessing). Adds:

• observation.images.image_dinov3 (float32[1024])
• observation.images.image_siglip2 (float32[768]) Removes:
• observation.images.image

2. Task-Text Embeddings (one row per unique task) build_task_text_embeddings.py with SentenceTransformer("google/embeddinggemma-300m") → data/task_text_embeddings.parquet + meta/task_text_embeddings_info.json.

3. Data Consolidation (this release) All per-episode Parquets were consolidated into N large Parquet shards in one data/ folder.

• The index meta/sharded_index.json records, for each episode, its normalized source identifier data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet, the destination shard path, and the (row_offset, num_rows) range inside that shard.
• This preserves original addressing while making Hub sync/clone/stream far faster and more reliable.

---

Metadata (Excerpts)

meta/task_text_embeddings_info.json

{
  "model": "google/embeddinggemma-300m",
  "dimension": 768,
  "normalized": true,
  "count": 5,
  "file": "task_text_embeddings.parquet"
}

meta/info.json (embedding-only + shards)

{
  "codebase_version": "v2.1-embeddings-sharded",
  "robot_type": "hello_stretch",
  "total_episodes": 135,
  "total_frames": 25016,
  "total_tasks": 5,
  "total_videos": 135,
  "total_chunks": 1,
  "chunks_size": 1000,
  "fps": 10,
  "splits": {
    "train": "0:135"
  },
  "data_path": "data/shard-{shard_id:05d}-of-{num_shards:05d}.parquet",
  "features": {
    "observation.state": {
      "dtype": "float32",
      "shape": [
        8
      ],
      "names": {
        "motors": [
          "x",
          "y",
          "z",
          "roll",
          "pitch",
          "yaw",
          "pad",
          "gripper"
        ]
      }
    },
    "action": {
      "dtype": "float32",
      "shape": [
        7
      ],
      "names": {
        "motors": [
          "x",
          "y",
          "z",
          "roll",
          "pitch",
          "yaw",
          "gripper"
        ]
      }
    },
    "timestamp": {
      "dtype": "float32",
      "shape": [
        1
      ],
      "names": null
    },
    "frame_index": {
      "dtype": "int64",
      "shape": [
        1
      ],
      "names": null
    },
    "episode_index": {
      "dtype": "int64",
      "shape": [
        1
      ],
      "names": null
    },
    "index": {
      "dtype": "int64",
      "shape": [
        1
      ],
      "names": null
    },
    "task_index": {
      "dtype": "int64",
      "shape": [
        1
      ],
      "names": null
    },
    "observation.images.image_dinov3": {
      "dtype": "float32",
      "shape": [
        1024
      ],
      "names": null
    },
    "observation.images.image_siglip2": {
      "dtype": "float32",
      "shape": [
        768
      ],
      "names": null
    }
  },
  "video_keys": [
    "observation.images.image"
  ],
  "num_shards": 64,
  "index_path": "meta/sharded_index.json"
}

---

Environment & Dependencies

Python ≥ 3.9 • PyTorch ≥ 2.1 • transformers • sentence-transformers • pyarrow • tqdm • decord (and optionally av)

---

Provenance, License, and Citation

• Source dataset: IPEC-COMMUNITY/cmu_stretch_lerobot
• License: apache-2.0 (inherits from the source)
• Encoders to cite:
  • facebook/dinov3-vitl16-pretrain-lvd1689m
  • google/siglip2-base-patch16-384
  • google/embeddinggemma-300m

---

Changelog

• v2.0-embeddings-sharded — Replaced video tensors with DINOv3 + SigLIP2 features; added EmbeddingGemma task-text embeddings; consolidated per-episode Parquets into N shards with a repo-local index; preserved original indexing/splits via normalized episode identifiers.