StanFu/Hide-and-Seek-v1

Hide-and-Seek Dataset

The Hide-and-Seek (HAS) dataset is a large-scale multimodal tactile dataset for tactile-only object recognition and retrieval in vision-denied environments.

This dataset accompanies the ICRA 2026 paper:

Tactile Hide and Seek: Bimanual Object Blind Search and Retrieval Via Tactile-Only Feedback

The dataset was collected on the Tactile Omnidirectional Mobile Manipulator (TOMM) using hand-mounted tactile skin sensors, wrist force-torque sensing, and robot proprioception. It supports learning-based tactile perception for object classification, weight classification, and bimanual verification.

Project page: https://tactile-hide-and-seek.github.io/

---

Dataset Summary

HAS contains 1,120,116 frame-wise tactile examples, approximately 1.12M frames, grouped into approximately 5.4k trajectories.

The dataset covers:

• 33 physical objects
• 1 negative class
• 61 object-weight categories
• 4 weight labels: none, light, medium, heavy
• Multimodal tactile observations from robot skin, force-torque sensing, and end-effector pose

The dataset is stored in Parquet format and provides official train / validation / test splits.

---

Data Splits

Split	Examples
Train	777,169
Validation	221,195
Test	121,752
Total	1,120,116

---

Features

Each row corresponds to one frame of synchronized tactile interaction data.

Feature	Shape / Type	Description
timestamp	float64	Frame timestamp
episode_index	int64	Episode / trajectory identifier
obs_point_cloud_skin_contact	88 × 6	Tactile contact point cloud from robot skin
mask_point_cloud_skin_contact	88	Boolean mask indicating valid tactile skin points
obs_pose_left	7	Left end-effector pose [x, y, z, qx, qy, qz, qw]
obs_skin_prox_left	44	Left-hand tactile skin proximity readings
obs_skin_force_left	44	Left-hand tactile skin force readings
obs_skin_dist_left	44	Left-hand calibrated distance readings
obs_dami_force_left	6	Aggregated DAMI force representation
obs_dami_prox_left	6	Aggregated DAMI proximity representation
obs_ft_left	6	Wrist force-torque wrench
classification_label	int64	Object-weight classification label

The tactile point cloud has shape 88 × 6, where the 88 points correspond to tactile skin cells across both robot hands. Each point contains spatial and sensor-derived channels.

---

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("StanFu/Hide-and-Seek-v1")

print(dataset)
print(dataset["train"][0].keys())

Example output:

DatasetDict({
    train: Dataset(...)
    validation: Dataset(...)
    test: Dataset(...)
})

Access one sample:

sample = dataset["train"][0]

point_cloud = sample["obs_point_cloud_skin_contact"]
mask = sample["mask_point_cloud_skin_contact"]
label = sample["classification_label"]

print(len(point_cloud))  # 88
print(label)

---

Task Description

The dataset supports tactile-only classification from robot interaction data.

Given a window of tactile observations, the learning task is to predict:

1. Object identity
2. Weight category

The ICRA 2026 paper uses a dual-head multimodal classifier that jointly predicts object class and weight class from synchronized tactile modalities.

---

Modalities

HAS includes five synchronized sensing streams:

1. Tactile point cloud Contact geometry and sensor intensities from distributed tactile skin cells.

2. Raw skin signals Per-cell proximity, force, and distance readings.

3. Virtual skin wrenches Aggregated proximity- and force-based wrench representations.

4. Proprioceptive pose End-effector pose from the robot kinematic state.

5. Wrist force-torque wrench Six-axis force / torque measurements from the wrist sensor.

---

Dataset Collection

The dataset was collected in a vision-denied tabletop workspace using TOMM, a bimanual robot equipped with hand-mounted tactile skin.

The data collection protocol includes:

• Single-hand pushing interactions for tactile object classification
• Bimanual grasping interactions for verification
• Randomized object placements
• Multiple object-weight variants
• Data collection across different days to reduce overfitting to a single sensor state

The official train / validation / test splits are provided for reproducible evaluation.

---

Baseline Results

The ICRA 2026 paper reports the following representative results on the HAS dataset and online robot trials:

Metric	Result
Object classification accuracy	91.1%
Weight classification accuracy	83.1%
Strict joint accuracy	79.6%
Online object + weight retrieval success	61.4%

The bimanual verification stage further corrects up to 17.6% of single-hand errors in online trials.

---

Label Mapping

The classification_label field encodes 61 object-weight categories as integer labels.

The full mapping is provided in label_mapping.yaml.

Each label follows the format:

<object_name>_<weight_class>

where the weight class is one of:

light, medium, heavy

The special label negative corresponds to non-target / negative examples.

---

Citation

If you use this dataset, please cite:

@inproceedings{fu2026tactilehideandseek,
  title     = {Tactile Hide and Seek: Bimanual Object Blind Search and Retrieval Via Tactile-Only Feedback},
  author    = {Fu, Xiangyu and Xing, Hao and Armleder, Simon and Shen, Wenlan and Wang, Fengyi and Guadarrama-Olvera, Julio Rogelio and Cheng, Gordon},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2026}
}

---

Contact

For questions, please contact:

Xiangyu Fu Technical University of Munich Project page: https://tactile-hide-and-seek.github.io/