KLHWONG/MIDFOV_EXPERIMENT

MIDFOV Experiment

Simulation demonstrations for Equivariant Visual-Tactile Diffusion Policy for Contact-Rich Manipulation (VISTA), accepted to the Conference on Robot Learning (CoRL 2026).

• Paper: https://arxiv.org/abs/2610.03333
• Project page: https://vista-paper.github.io/
• Code: https://github.com/Kenn3o3/Vista
• Real-robot code: https://github.com/Kenn3o3/vt_franka
• Same release on ModelScope: https://modelscope.cn/datasets/kenn3o3/MIDFOV_EXPERIMENT

Eight contact-rich tasks, 100 successful episodes each. Episodes are stored as HDF5, with a metadata.json per task and rollout videos.

Task	Episodes
grasp_classify	100
insert_HDMI	100
insert_hole	100
insert_tube	100
lift_bottle	100
lift_can	100
pull_out_key	100
put_bottle_in_shelf	100

Layout

<task>/hdf5/<episode_id>.hdf5
<task>/metadata.json
videos/<task>/<episode_id>_success.mp4

The training code reads this directory directly. It does not use load_dataset.

Download

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="KLHWONG/MIDFOV_EXPERIMENT",
    repo_type="dataset",
    local_dir="data/MIDFOV_EXPERIMENT",
)

Or from the Vista repository:

hf download KLHWONG/MIDFOV_EXPERIMENT --repo-type dataset --local-dir data/MIDFOV_EXPERIMENT

Citation

@inproceedings{wong2026vista,
  title={Equivariant Visual-Tactile Diffusion Policy for Contact-Rich Manipulation},
  author={Wong, Lik Hang Kenny and Ma, Yiyao and Wei, Xiu-Shen and Tan, Zelong and Song, Zhuheng and Xie, Dongsheng and Chen, Kai and Dou, Qi},
  booktitle={Conference on Robot Learning},
  year={2026},
}