README.md

---
dataset_info:
  features:
    - name: observation.images.head_rgb_image
      dtype: video
    - name: observation.images.hand_rgb_image
      dtype: video
    - name: observation.images.hand_depth_image
      dtype: video
    - name: observation.state
      dtype: sequence
    - name: action
      dtype: sequence
  splits:
    - name: train
      num_examples: 80
license: other
task_categories:
  - robotics
tags:
  - robotics
  - imitation-learning
  - vision-language-action
  - hsr
  - mujoco
  - lerobot
---

# HSR Tomato Can Box RandomRect RealMatch Mix80

This dataset is a LeRobot-style HSR mobile-manipulation dataset used by the
Evo-Mibo demo training code.

It contains 80 demonstrations for the tomato-can-box task with:

- head RGB video
- hand RGB video
- hand inverse-depth video
- 15D robot state
- compact 2D scene slots, forming a 27D state vector
- 9D action
- language instruction

The task instruction is:

```text
Pick up the can and move it to the box.
```

The matching code and training scripts are expected to use:

```text
dataset/config_demo_hsr_tomato_randomrect_realmatch_mix80_state27.yaml
```

Example download:

```bash
hf download Maxmibo/HSR-TomatoCanBoxRandomRectRealMatchMix80 \
  --repo-type dataset \
  --local-dir data/HSR_TomatoCanBoxRandomRectRealMatchMix80_RGB15_HandInvDepth005to200_State27_BinaryGripper025_InitialScene_20260621_20260720_lerobot_v21
```