WISER testbed data

Data for Grounded World Model: Latent Planning with Language Goals.

Code and setup · GWM checkpoints

WISER is a controlled manipulation testbed in ManiSkill for studying semantic generalization. It provides 288 training tasks and 288 held-out test tasks with unseen instructions and visual signals. The LeRobot datasets were collected with LeRobot 0.4.3 and use the LeRobotDataset v3.0 format.

Contents

Folder	Purpose	Episodes	Tasks	Frames
merged_train/	Training demonstrations	1,728	288	207,360
merged_test/	Held-out test demonstrations	288	288	34,560
no_noise_demo_1_round/config_{0..23}_train/lerobot_data/	Recorded future observations for GT-MPC, training tasks	288 total	288 total	34,560 total
no_noise_demo_1_round/config_{0..23}_test/lerobot_data/	Recorded future observations for GT-MPC, test tasks	288 total	288 total	34,560 total

Each GT-MPC configuration folder contains 12 episodes for 12 tasks. Episode metrics are stored alongside its lerobot_data/ folder. rlds_train/ provides the training data converted to RLDS/TFRecord for compatible VLA training workflows; it is an alternative format, not an additional held-out split.

LeRobot folders contain data/ Parquet files, meta/ episode and task metadata, and videos/ observations at 20 Hz. The protocol split is determined by the folder name; the local LeRobot metadata in both merged folders calls its episode range train.

Download and use

Use direct downloads to preserve the LeRobot directory layout:

hf download Shady0057/WISER --repo-type dataset \
    --include 'merged_train/**' 'merged_test/**' 'no_noise_demo_1_round/**' \
    --local-dir wiser_dataset

Download the RLDS conversion separately if your workflow requires it:

hf download Shady0057/WISER --repo-type dataset \
    --include 'rlds_train/**' --local-dir wiser_dataset

Point the repository's training and evaluation entrypoints to the downloaded folders. Learned GWM planning uses the RGB-free action skills in the code repository's gwm_skills directory and predicts future embeddings; GT-MPC uses the recorded future observations here.

The generic Hub viewer is disabled because these folders combine LeRobot episode tables, metadata and video files. Follow the direct-download workflow above to use the original data.

Citation

@misc{li2026groundedworldmodellatentplanning,
      title={Grounded World Model: Latent Planning with Language Goals},
      author={Quanyi Li and Lan Feng and Haonan Zhang and Wuyang Li and Letian Wang and Alexandre Alahi and Harold Soh},
      year={2026},
      eprint={2604.11751},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2604.11751},
}