README.md

---
pretty_name: DROID SAM 3.1 Segmentation Masks
task_categories:
- image-segmentation
tags:
- robotics
- rlds
- droid
- sam3
configs:
- config_name: default
  data_files:
  - split: train
    path: data/*.parquet
---

# DROID SAM 3.1 Segmentation Masks

This dataset is a mask-only sidecar generated from the original
`droid_101/0.0.1` RLDS release. It does not redistribute DROID images or
actions. Its `episode_index` follows the RLDS episode order.

The same episodes appear in `lerobot/droid_1.0.1`, but LeRobot stores them in a
different episode order. Therefore, mask `episode_index` and LeRobot
`episode_index` must not be joined directly. Use the mapping file described
below to associate these masks with LeRobot episodes.

## Prompts

- `exterior_image_1_left`: `robot manipulator`
- `exterior_image_2_left`: `robot manipulator`
- `wrist_image_left`: `end-effector`

Masks are per-step semantic unions encoded as compressed COCO RLE dictionaries
with `size` and ASCII `counts`. Episodes were processed in independent chunks
of `800` frames (`0` means a full episode).

Source TFDS subsplits: `201`

SAM source revision: `46957e47805eaa273f4aa7bbbd25a88bca9108ce`

Weight set: `gripper_universal`

SAM checkpoint revision: `daa63191845a41281374e725f4c9e51c7a824460`

Gripper detector repository: `sazirarrwth99/sam3.1-gripper-universal`

Gripper detector revision: `be4e08e989a673511b11b7214fb135639dc5132a`

## LeRobot to RLDS episode mapping

[`lerobot_to_rlds_episode_mapping.parquet`](lerobot_to_rlds_episode_mapping.parquet)
provides the episode correspondence between LeRobot and RLDS. The mapping
contains 95,658 rows, one per LeRobot `hf_episode_index`. The RLDS shard,
record, and source identity are included for additional verification.

**Important:** `rlds_shard_major_index` and this dataset's `episode_index`
column are both positional counters, not stable identifiers — do not use
either to index directly into `data/*.parquet`. Always join on `episode_id`
(this dataset's column) against `rlds_episode_id` (the mapping's column),
which is content-addressed and stable across row-order changes.

LeRobot-side fields include `hf_episode_index`, `hf_length`, `hf_data_path`,
`hf_dataset_from_index`, and `hf_dataset_to_index`. RLDS-side fields include
`rlds_shard_index`, `rlds_record_index`, `rlds_shard_major_index`,
`rlds_tfrecord_path`, `rlds_byte_offset`, `rlds_file_path`,
`rlds_recording_folderpath`, and `rlds_episode_id`.

All 95,658 episodes were mapped with zero unmatched or unresolved episodes.
Matches were verified using exact equality of the complete float32
`action.joint_position` sequence. Of these, 95,466 mappings are unique. The
remaining 192 rows form 96 exact duplicate-equivalence groups: the RLDS source
contains indistinguishable copies of the same recording under aliased source
identities. For these rows, the physical record is paired deterministically;
copy-level provenance cannot be recovered from identical episode content. Check
`mapping_uniqueness`, `rlds_candidate_count`, and the `candidate_rlds_*` columns
when this distinction matters.

Example lookup in either direction:

```python
import pyarrow.compute as pc
import pyarrow.parquet as pq
from datasets import load_dataset

mapping = pq.read_table("lerobot_to_rlds_episode_mapping.parquet")
masks = load_dataset("EpicPinkPenguin/droid_dataset_segmentation_mask", split="train")
episode_id_to_row = {episode_id: row for row, episode_id in enumerate(masks["episode_id"])}

# LeRobot episode -> mask row
match = mapping.filter(pc.equal(mapping["hf_episode_index"], 12345))
rlds_episode_id = match["rlds_episode_id"][0].as_py()
mask_row = episode_id_to_row[rlds_episode_id]

# Mask row -> LeRobot episode
episode_id = masks[mask_row]["episode_id"]
match = mapping.filter(pc.equal(mapping["rlds_episode_id"], episode_id))
lerobot_episode_index = match["hf_episode_index"][0].as_py()
```

SHA-256: `7548f287fb4f2d62c642e4e89a8b7a5ac50d278d18b1d2ff0b042e4c7309eb5a`

Generated with the DROID segmentation annotation pipeline. The original DROID
dataset and SAM checkpoint remain subject to their own licenses and terms.