Rongxuan-Zhou/pusht_lewm_fr3

PushT-LeWM-FR3 — Real-Robot Franka PushT for World-Model Planning

Real-robot PushT demonstrations collected on a Franka Research 3 (FR3) by Meta Quest 3 teleoperation, stored in the LeWM flat HDF5 format used by the LeWorldModel / PRISM world-model trainer (HDF5Dataset reads raw pixels — no video decode). This is the real-robot PushT dataset behind the hardware results in the PRISM paper (Sec. 4.4).

• Paper: PRISM: PRior-guided Imagination Sampling in world Models — arXiv:2606.07974
• Project page: https://YuhaiW.github.io/PRISM_web/
• Code: https://github.com/YuhaiW/prism-jepa
• Robot: Franka Research 3, planar PushT with a rigid push-rod
• Task: push a T-shaped block to a target pose on a planar wooden table
• Camera: single third-person Intel RealSense D455 (agent_view, top-down-ish)
• Teleop: Meta Quest 3 (WebXR)

Dataset at a glance

Episodes	500
Frames (transitions)	111,531
Control rate	10 Hz
Episode length	mean 223 frames (~22.3 s), min 93, max 703
Image	224×224×3 uint8
Action	6-D end-effector delta
Proprioception	7-D end-effector pose
Single file	pusht_lewm_fr3.h5 (~10.4 GB)

Schema — pusht_lewm_fr3.h5 (flat HDF5)

N = 111,531 frames, E = 500 episodes. Frames are concatenated across episodes; use episode_idx / ep_offset / ep_len to recover episode boundaries.

key	shape	dtype	meaning
pixels	(N, 224, 224, 3)	uint8	D455 agent_view, resized. Channel order: BGR
action	(N, 6)	float32	consecutive EE delta [dx, dy, dz, drx, dry, drz] @10 Hz (m, rad)
proprio	(N, 7)	float32	EE pose [x, y, z, qw, qx, qy, qz] in robot base frame (m, unit quat)
state	(N, 7)	float32	= proprio (no external object-pose tracking)
ep_len	(E,)	int32	per-episode frame count
ep_offset	(E,)	int64	per-episode start index into the flat arrays
episode_idx	(N,)	int64	episode id for each frame
step_idx	(N,)	int64	step index within the episode

Root attributes: fps=10.0, task="pusht", robot_type="franka_panda" (FR3 uses the Panda kinematic model in libfranka), action_names="dx,dy,dz,drx,dry,drz", proprio_names="ee_x,ee_y,ee_z,ee_qw,ee_qx,ee_qy,ee_qz".

Value ranges (observed)

• proprio xyz (m): x ∈ [0.22, 0.85], y ∈ [-0.51, 0.49], z ∈ [0.085, 0.116] (the rod is held in a thin planar band ~z≈0.10 m above the table; the small z spread reflects the planar lock).
• action (m / step): dx ∈ [-0.045, 0.052], dy ∈ [-0.049, 0.049], dz ∈ [-0.022, 0.012]; rotational deltas are near-zero (planar task).

Intended use

Designed for embedding-space world models (JEPA / LeWM) and sampling-based planning (MPPI / CEM), as used in PRISM. Typical consumers:

• Train a JEPA latent world model from pixels + action.
• Train an action-intuition / behavior-cloning head on (pixels, action) for prior-guided sampling.
• Goal-conditioned eval by relabeling a future frame as goal via episode_idx / step_idx.

PRISM planner settings (from the paper, for reproducibility): action block / frame-skip 5, planning horizon H=5, MPPI J=30 iterations.

Loading

import h5py
from huggingface_hub import hf_hub_download

path = hf_hub_download("Rongxuan-Zhou/pusht_lewm_fr3",
                       "pusht_lewm_fr3.h5", repo_type="dataset")
with h5py.File(path, "r") as f:
    pixels      = f["pixels"]          # (N,224,224,3) uint8, BGR — index lazily, do not load all
    action      = f["action"][:]       # (N,6) float32
    proprio     = f["proprio"][:]      # (N,7) float32
    ep_len      = f["ep_len"][:]       # (500,)
    ep_offset   = f["ep_offset"][:]    # (500,)

    # iterate one episode
    e = 0
    s, n = int(ep_offset[e]), int(ep_len[e])
    ep_pixels  = pixels[s:s+n]         # (n,224,224,3)
    ep_action  = action[s:s+n]

pixels is BGR (as captured by OpenCV). If your model expects RGB, convert with img[..., ::-1]. Keep the convention consistent between training and deployment.

Suggested split

No official split is shipped. Split by episode (not by frame) to avoid leakage, e.g. hold out a random 10% of the 500 episode ids for validation.

Collection notes & curation

• Teleop: WebXR / Meta Quest 3, planar lock (translation in x–y; z and flange orientation held by the controller), Cartesian impedance servo at 1 kHz, targets at ~10 Hz.
• Action derivation: from logged EE pose, resampled to a uniform 10 Hz grid by wall-clock timestamps (position lerp + quaternion slerp), ee_ok-filtered, then consecutive delta. Crash-safe writer (per-episode checkpoint + flush).
• Exposure: auto-exposure converged then frozen with a neutral color profile, so the white rod stays distinct from the wooden table across demos.
• Curation: live-curated during collection (jittery / static / failed demos discarded).

Limitations

• No wrist/egocentric view and no depth (single third-person RGB).
• Demonstrations are near-expert teleoperation, not optimal — suitable as a world-model / prior training set, not as a verified optimal-control benchmark.
• Goal pose not stored as a field (relabel from future frames).
• Mild visual domain drift across collection sessions (lighting, arm posture).

License

Released under the MIT License. The data contains no human subjects or personally identifiable information (robot end-effector and a tabletop block only).

Citation

If you use this dataset, please cite PRISM:

@misc{wang2026prismpriorguidedimaginationsampling,
      title={PRISM: PRior-guided Imagination Sampling in world Models},
      author={Yuhai Wang and Jiawei Xia and Rongxuan Zhou and Xiao Hu and Yongliang Shi and Jing Du and Yang Ye},
      year={2026},
      eprint={2606.07974},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2606.07974},
}