Datasets:

---

nvidia

/

PhysicalAI-Robotics-GR00T-X-Embodiment-Sim

like 226

Follow

NVIDIA 56.7k

Hi, thanks for releasing the PhysicalAI-Robotics-GR00T-X-Embodiment-Sim dataset.

I am trying to understand whether the released dataset can be replayed in the original simulator, not only used as offline RGB/state/action trajectories.

From the released LeRobot-style dataset, I can find files such as:

• data/chunk-000/episode_*.parquet
• videos/chunk-000/.../episode_*.mp4
• meta/info.json
• meta/modality.json
• meta/episodes.jsonl
• meta/tasks.jsonl

The dataset contains observation.images.*, observation.state, and action. However, I could not find per-episode simulator replay data, such as:

• full simulator states for each frame
• initial simulator state
• per-episode scene configuration / randomization metadata
• object poses or object states
• camera intrinsics / extrinsics used for rendering
• scene files or episode-level references needed to reconstruct the exact trajectory

My understanding is that the simulator assets / environments may be available for evaluation, but exact playback of a recorded demonstration still requires per-episode simulator states or equivalent replay metadata. For example, in RoboCasa-style playback, a replayable dataset usually needs something like per-episode states, model.xml, and ep_meta.

Could you clarify:

1. Is observation.state intended to be sufficient for simulator playback, or is it only the robot proprioceptive state?
2. Are the full simulator states or equivalent replay metadata available somewhere for this dataset?
3. Is there an official script or recommended pipeline to replay a released episode in the simulator?
4. If exact playback is not supported from the released files, is that by design, or are there plans to release the missing per-episode replay information?

My goal is to reconstruct the recorded trajectories in the simulator as accurately as possible. Any guidance on the intended replay/evaluation workflow would be very helpful.

Thanks!