ExylosAi/object-sorting-cross-embodiment-rich-modality-sample

Cross-Embodiment Object Sorting — Rich-Modality 20-Episode Inspection Sample

20 episodes total: 10 Franka Panda + 10 WidowXAI. A compact cross-embodiment inspection release for picking up an instructed object and placing it into an instructed target box. Each robot keeps its native LeRobot v2.1 state/action schema in a separate Viewer config. Five synchronized RGB views, metric depth and instance segmentation for every camera, robot state/action, end-effector trajectories, gripper semantics, and QA metadata are included.

Use it / Skip it

Use it for

• Inspecting cross-embodiment manipulation schemas before working with larger releases.
• Testing multi-view RGB, depth, segmentation, robot-state, and end-effector preprocessing.
• Comparing the same sorting task across Franka Panda and WidowXAI.
• Validating loaders, visualization tools, and embodiment adapters.

Skip it if you need

• A complete policy-training corpus. Twenty episodes are intended for inspection and pipeline validation.
• A single shared state/action vector across robots.
• Force/torque, tactile, contact, point-cloud, or audio streams.
• Human-reviewed natural-language annotations; the included labels were reconstructed automatically from the original recordings.

At a glance

Episodes / frames	20 / 39,976 (≈22.21 min at 30 FPS)
Robot split	10 Franka Panda + 10 WidowXAI
Task	Episode-specific object sorting into Box A, B, or C
State/action	Panda 9-D; WidowXAI 8-D; separate robot configs
RGB	Five synchronized 1280×960 H.264 views, no audio
Depth	Metric float32 depth for all five cameras
Segmentation	Instance segmentation for all five cameras, shared 41-class label map
End-effector	Local/world TCP pose, velocity, achieved next-frame Cartesian action
Format	Two robot-specific LeRobot v2.1 subsets plus four Hugging Face Viewer configs
Repository footprint	≈65.93 GB
License	Not specified in the source repository

All five included RGB views already contain metric depth and instance segmentation. Additional synthetic modalities can be supplied on request.

Load it

Select the embodiment explicitly:

from datasets import load_dataset

repo = "ExylosAi/object-sorting-cross-embodiment-rich-modality-sample"

panda = load_dataset(repo, name="panda", split="train")
widowxai = load_dataset(repo, name="widowxai", split="train")
episodes = load_dataset(repo, name="episodes", split="train")
videos = load_dataset(repo, name="videos", split="train")

The robot frame schemas remain separate. Do not concatenate their state/action vectors without an explicit embodiment adapter.

Canonical nested Parquet is stored under each robot root:

from huggingface_hub import hf_hub_download
import pyarrow.parquet as pq

path = hf_hub_download(
    "ExylosAi/object-sorting-cross-embodiment-rich-modality-sample",
    "robots/panda/data/chunk-000/episode_000000.parquet",
    repo_type="dataset",
)
episode = pq.read_table(path)
print(episode.schema)

Episode inventory

Robot	Episodes	Frames	Duration
Franka Panda	10	24,739	≈13.74 min
WidowXAI	10	15,237	≈8.47 min
Total	20	39,976	≈22.21 min

Use the episodes config for episode-level provenance, duration, robot type, source QA status, and source-recording metadata. Use videos to browse all 100 RGB streams.

Modalities

• RGB: front_cam, top_cam, wrist_cam, right_cam_45, and left_cam_45.
• Depth: packed float32 NPZ in meters for every camera.
• Segmentation: lossless FFV1 MKV with a shared 41-class label map.
• Robot state/action: native embodiment-specific joint and gripper vectors, without cross-robot padding.
• End-effector: local and world TCP poses, linear velocity, and Cartesian state/action.
• Gripper: open scale, mode, and activity annotations.
• QA: D1–D7 scores, composite score, confidence, grade, frozen-frame count, and reconstructed task-outcome fields.

Provenance

sample_manifest.json preserves the mapping between source episode IDs, robot-specific local IDs, source recordings, and URDFs. Media files are server-side copies of the immutable source payload; episode paths are remapped only to create independent robot subsets.

File layout

robots/panda/        meta · data · videos · annotations for 10 Panda episodes
robots/widowxai/     meta · data · videos · annotations for 10 WidowXAI episodes
viewer_data/         separate flattened Panda and WidowXAI frame configs
viewer_index/        combined 20-row episode and provenance config
viewer_videos/       combined 100-row RGB Video-feature config
assets/              synchronized front-camera RGB/depth/segmentation preview
sample_manifest.json · annotations.json

Notes and limitations

• This is synthetic data intended for inspection and pipeline validation, not a statistically representative benchmark.
• Panda and WidowXAI use different native joint schemas.
• Joint-space and Cartesian actions represent achieved next-frame state, not recorded controller commands.
• Per-frame 6-DoF object poses are not included.
• Instruction, target-object, target-box, outcome, phase, and subtask fields were reconstructed from the original recordings.
• The reconstructed annotations have not been independently reviewed by a human.
• No standalone license file or explicit license terms were present in the source repository. Do not assume redistribution rights.

Citation

If you use this sample, cite this repository together with its exact commit.

Annotation correction

The per-episode instructions, target objects, target boxes, outcomes, phase annotations, and quality scores were reconstructed from the original recordings. Episode alignment was verified using exact frame counts, timestamps, frame indices, and end-effector trajectories. WidowXAI D7 scores use the six actuated joints and URDF position/velocity limits.