yilin-wu/ViTaL-datasets

ViTaL datasets

Paper: Inference-time Policy Steering via Vision and Touch · Project website · Code

Robot trajectories for pipette liquid transfer, peg insertion, and whiteboard wiping. Each episode includes two RGB views, optical tactile images, precomputed visual/tactile embeddings, actions, and robot state.

Task	Episodes	Train / test	Frames	Compressed GB
pipette	320	270 / 50	153,498	361.02
insertion	330	280 / 50	85,395	200.75
wiping	330	280 / 50	63,738	154.74

The split uses lexicographically sorted original trajectory IDs, shuffled with seed 42, with the first 50 trajectories held out per task. Original IDs, numeric dtypes, array values, and normalization statistics are preserved. The HDF5 datasets use lossless gzip compression with chunk checksums.

Download

python -m pip install "huggingface_hub>=0.36,<1"
hf download yilin-wu/ViTaL-datasets --repo-type dataset --include "pipette/*" --local-dir datasets/vital

Replace pipette/* with insertion/* or wiping/* for one task, or omit --include for all three. Authenticate with hf auth login first if this repository is private.

pipette/                     # also insertion/ and wiping/
  episodes/trajectory_*.hdf5
  normalization.json
  splits.json
  manifest.jsonl
  metadata.json

Use with the released world model

After following the installation instructions, run these commands from vital/latent-tactile:

export TACTILE_DATASET=/absolute/path/to/datasets/vital/pipette/episodes
export TACTILE_NORM_STATS=/absolute/path/to/datasets/vital/pipette/normalization.json
bash scripts/train_liquid.sh --no_consolidated

export WM_CHECKPOINT=/path/to/wm_kstep_concat_h1_p1_fs8_iter67000.pth
export TACTILE_DECODERS=/path/to/liquid_decoders
bash scripts/evaluate_liquid.sh --no_consolidated

These commands use the liquid recipe and checkpoint. Insertion and wiping require their own trained checkpoints and decoder weights. World-model, decoder, AnyTouch, and diffusion-policy weights are not included in this dataset.

Episode schema

Each file has a data group. T denotes the action count. R denotes a raw observation count of either T or T+1, depending on the field and episode. The group retains a controller-configuration JSON attribute when present in the source (100 pipette, 100 insertion, and 50 wiping episodes), and includes the original trajectory ID as an attribute.

Field	Shape	Stored dtype
actions	(T, 7)	float64
states	(T, 8)	float64
camera_0, camera_1	(R, 224, 224, 3)	float32
camera_2	(R, H, W, 3)	uint8
cam_0_patch_embd, cam_1_patch_embd	(T, 196, 768)	float32
cam_tactile_patch_embd	(T, 392, 512)	float32
cam_0_cls_embd, cam_1_cls_embd	(T, 768)	float32
cam_tactile_cls_embd	(T, 512)	float32
ee_states	(R, 16)	float64
gripper_states	(R,)	float64
joint_states	(R, 7)	float64

Tactile images are 240×320 for pipette/wiping and 256×256 for insertion. The released world-model loader takes the first 196 tactile patch tokens. The full 392 stored tokens remain available in this release.

Some raw camera and auxiliary robot-state arrays retain an extra terminal observation: 56 pipette, 43 insertion, and 29 wiping episodes have at least one such array. Actions, model states, and embeddings all have length T. The released loaders use the first T raw samples; no terminal observation has been discarded or synthesized. Per-array shapes are in the manifest.

Actions are stored before normalization. Use the accompanying mean_acs and std_acs for all seven action dimensions; the active world-model recipe does not normalize the eight-dimensional state.

These files do not provide per-episode natural-language instruction or success labels. The task directory identifies the task family. The exported trajectories are world-model training data, not a benchmark of the released controller's success rate. No temporal resampling or numeric downcasting was applied.

Integrity and provenance

manifest.jsonl records file SHA-256 checksums and per-array value checksums. Every exported array was read back and compared by hash with its source. metadata.json identifies the original consolidated source filename and the normalization checksum. Source datasets were opened read-only.

License

Creative Commons Attribution 4.0 International (CC BY 4.0).

Citation

@misc{wu2026vital,
  title={Inference-time Policy Steering via Vision and Touch},
  author={Yilin Wu and Zilin Si and Zeynep Temel and Oliver Kroemer and Andrea Bajcsy},
  year={2026},
  eprint={2606.14981},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2606.14981}
}