windjj/datumwell-embodiedtrace

Datumwell EmbodiedTrace

See the task. Explore the subgoal. Read the action.

A 100-episode preview connecting visual observations, task-oriented annotations, and structured robot actions. Explore a concrete data representation for embodied-agent development, then assess how it fits your tasks and action interface.

Explore the episodes → · Project overview · Discuss a pilot

What's in the sample

Included	Count
Episodes	100
Structured EEF decisions	4,081
Generated subgoal segments	1,288
Source timesteps	32,212
Camera views	3
Episodes with instructions	77
Episodes without instructions	23
Recorded successful / failed episodes	81 / 19

Each JSONL row is one decision, rather than one episode. Generated task-family groupings include 52 transfers, 8 multi-step tasks, 4 surface moves, 4 articulated-object tasks, 4 presses, 2 wipes, 1 rotation, and 1 scoop. Another 24 episodes have unresolved generated task-family categorization; this is distinct from the 23 missing instructions. These groupings are generated labels, not audited task taxonomy. See coverage.json and the episode index.

Inspect one decision in five minutes

Install the lightweight dependencies for this example:

pip install datasets huggingface_hub

from datasets import load_dataset

sample = load_dataset("windjj/datumwell-embodiedtrace", split="train", streaming=True)
row = next(iter(sample))
print(row["task"])
print(row["generated_annotation"]["subgoal"]["title"])
print(row["action"])

This loads structured records. Images are portable path references; they are not automatically downloaded or decoded by this call. To inspect without the datasets library, download the repository's data/train.jsonl and use the standard-library loader.

Abridged example

The following excerpt comes from the first decision. It omits history, rationale evidence, and later feedback for readability; examples/decision.json contains the complete presentation excerpt, while data/train.jsonl contains full decision rows.

{
  "task": "Put the marker in the pot",
  "generated_subgoal": "Approach / align with marker",
  "action": {
    "name": "move_eef",
    "arguments": {
      "frame": "robot_base",
      "target_pose": {
        "position_m": [0.432365, 0.109185, 0.480900],
        "quaternion_xyzw": [-0.988317, -0.034010, -0.117116, 0.091413]
      },
      "gripper_closure": 0.0,
      "max_steps": 15
    }
  }
}

Numbers above are rounded for display. Full rows preserve recorded numeric precision.

Optional media

Images archive: 286.1 MB, containing 12,543 JPEG keyframe observations. Videos archive: 99.7 MB, containing 100 three-view MP4 previews. Media is optional when inspecting the action schema.

from pathlib import Path
import importlib.util
from huggingface_hub import snapshot_download

root = Path(snapshot_download(
    repo_id="windjj/datumwell-embodiedtrace",
    repo_type="dataset",
    allow_patterns=["data/*", "media/images.zip", "examples/loader.py"]
))
spec = importlib.util.spec_from_file_location("embodiedtrace_loader", root / "examples/loader.py")
loader = importlib.util.module_from_spec(spec)
spec.loader.exec_module(loader)
loader.extract_media(root, "images")  # writes images/<episode_id>/<frame>_<camera>.jpg
# To inspect videos, download media/videos.zip separately first; then:
# loader.extract_media(root, "videos")  # writes videos/<episode_id>.mp4
row = next(loader.iter_decisions(root, "train"))
image_path = root / row["observation"]["images"][0]
print(image_path)

For repeatable work, pass a repository commit SHA as revision to load_dataset and snapshot_download. ZIP extraction writes media into the local snapshot directory and requires approximately the archive sizes again in free disk space.

Record structure

Field	Meaning
episode_id, decision_index, split	Episode identity, decision order, and split
task, instruction_available	Available task instruction; null when unavailable
observation	Current pose, gripper state, frame index, and three image paths
history	Past actions and recorded feedback
generated_annotation	Generated subgoal, concise rationale, action description, evidence, and provenance
action	Structured move_eef command target
feedback, progress_annotation	Later recorded observation and generated progress description
provenance	Frame alignment and derived-label status

Action positions use metres in the robot-base frame. Orientation uses unit quaternion XYZW. Gripper closure is 0=open, 1=closed. The source embodiment is a Franka Panda with a Robotiq gripper; coordinate targets are not directly portable to another robot without an appropriate action adapter. See action.schema.json.

Recorded commanded poses are represented as sparse EEF targets, with Euler XYZ rotations converted to quaternion XYZW. These derived macro-actions were not separately executed. Preview videos use three synchronized camera views at nominal 15 FPS; they are compressed inspection media, not sensor-timestamp evidence.

Evaluate fit

Use this sample to inspect representation compatibility, annotation needs, and task coverage. Prepare a pilot-scoping brief or download the checklist to share internally. A requirements discussion can scope a potential pilot around your tasks, observations, and action format; delivery scope and evaluation criteria would be agreed separately.

Generated labels are suggestions for inspection; action labels and later recorded feedback remain separate. Rationales are generated weak supervision, not demonstrator thoughts or verified reasoning. Object contact, grasp retention, and subgoal completion are unverified. Visual scene review was sampled, not dense at every decision.

For training, use current/past observations as inputs and generated annotations/actions as output labels. Keep later feedback and progress out of decision inputs. Some rationale text references the labeled action; do not treat it as independently observed pre-action reasoning. Review semantic labels before reasoning supervision.

The 80/10/10 train/validation/test split holds out episode IDs only. Scenes, operators, tasks, and recording dates are not held out. This is a representation demo, not a generalization benchmark. No model improvement or robot execution results are claimed.

Provenance and license

Original recordings and commands come from the publicly released 100-episode DROID sample, contributed by Khazatsky et al. and the DROID collaboration. Datumwell adds structured action labels, generated annotations, packaging, and presentation. This is a derived release, not a proprietary recording collection; no upstream endorsement is implied.

Data is released under CC BY 4.0, retaining source attribution and indicating changes. The loader code is MIT licensed. See NOTICE, LICENSE, CODE_LICENSE, the original project, and the source paper. Version: 1.0, released October 13, 2026. File fingerprints are in checksums.json.

@misc{datumwell_embodiedtrace_2026,
  title = {Datumwell EmbodiedTrace: A Structured Embodied-Agent Data Preview},
  author = {Datumwell},
  year = {2026},
  url = {https://huggingface.co/datasets/windjj/datumwell-embodiedtrace}
}
@inproceedings{khazatsky2024droid,
  title = {DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset},
  author = {Khazatsky, Alexander and Pertsch, Karl and Nair, Suraj and others},
  booktitle = {Robotics: Science and Systems},
  year = {2024}
}