README.md

---
license: apache-2.0
task_categories:
- robotics
tags:
- teleoperation
- dual-arm
- mcap
- TeleXperience
- DualArxR5a
- t-shirt-folding
---

# FoldingTShirt_DualArxR5a_Samples

<video src="https://huggingface.co/datasets/io-intelligence/FoldingTShirt_DualArxR5a_Samples/resolve/main/telexperience_preview.mp4" autoplay muted loop playsinline width="100%"></video>

100 real-robot teleoperation episodes for **“Fold the T-shirt on the table.”** on a **DualArxR5a** dual-arm robot. Format: raw [MCAP](https://mcap.dev/) (ROS 2 / rosbag2).

## Source

Collected with **[TeleXperience](https://io-ai.tech/en/telexperience/)**, IO-AI’s product for real-robot teleoperation and data collection. An operator drives the robot; TeleXperience writes time-aligned RGB, joint commands, joint states, gripper targets, and end-effector poses to MCAP.

Product page: https://io-ai.tech/en/telexperience/

**Task:** Fold the T-shirt on the table.  
**Robot:** DualArxR5a (bimanual, parallel-jaw grippers)  
**Language prompt:** `Fold the T-shirt on the table.`

## Repository layout

```text
FoldingTShirt_DualArxR5a_Samples/
  README.md
  telexperience_preview.mp4
  manifest.json
  data/
    <episode>.mcap
    <episode>.metadata.yaml
```

Episode file names keep the original TeleXperience recording id, for example:

`livingroom_T-shirt_folding_DualArxR5a_251110_145215_0.mcap`

Each `.metadata.yaml` is the rosbag2 inventory written at capture time (duration, topic list, message counts). Per-episode stats are also in `manifest.json`.

## MCAP data structure

Every episode is a self-contained MCAP with the same topic set.

| Topic | Type | Role | Typical rate |
| --- | --- | --- | --- |
| `/camera_high/color/image_raw/compressed` | `sensor_msgs/CompressedImage` | High / head RGB | ~30 Hz |
| `/camera_low/color/image_raw/compressed` | `sensor_msgs/CompressedImage` | Low / front RGB | ~30 Hz |
| `/camera_left_wrist/color/image_raw/compressed` | `sensor_msgs/CompressedImage` | Left-arm wrist RGB | ~30 Hz |
| `/camera_right_wrist/color/image_raw/compressed` | `sensor_msgs/CompressedImage` | Right-arm wrist RGB | ~30 Hz |
| `io_teleop/joint_cmd` | `sensor_msgs/JointState` | Teleop joint command (action) | ~100 Hz |
| `io_teleop/joint_states` | `sensor_msgs/JointState` | Robot joint state (proprioception) | ~100 Hz |
| `io_teleop/target_gripper_status` | `sensor_msgs/JointState` | Gripper command (`left_gripper`, `right_gripper`) | ~100 Hz |
| `io_teleop/target_ee_poses` | `geometry_msgs/PoseArray` | Target end-effector poses | ~100 Hz |
| `io_teleop/robot_info` | `std_msgs/String` | Robot / session metadata (`DualArxR5a`) | 1 message |

### Joint command and state (1:1)

`io_teleop/joint_cmd` and `io_teleop/joint_states` use the **same 14 names, same order, same units**:

```text
left_joint1, left_joint2, left_joint3, left_joint4, left_joint5, left_joint6, left_joint7,
right_joint1, right_joint2, right_joint3, right_joint4, right_joint5, right_joint6, right_joint7
```

- `*_joint1` … `*_joint6`: arm joints in **radians**
- `*_joint7`: gripper in **0 / 1** (open / closed), taken from `target_gripper_status`
- Look up by name; do not assume a converter-specific left/right packing

`target_gripper_status.names` is `left_gripper, right_gripper`, also **0 / 1**.

### Camera positions

All four streams are **640×480 JPEG**, synchronized at about 30 Hz.

| Topic | Mount | What you see |
| --- | --- | --- |
| `/camera_high/color/image_raw/compressed` | **High / head**, looking down at the workspace | Full white table, the T-shirt, both grippers, room background. Most stable view. |
| `/camera_low/color/image_raw/compressed` | **Low / front (chest)**, closer to the table | Same scene from a lower, nearer angle. Grippers occupy more of the bottom of the image. |
| `/camera_left_wrist/color/image_raw/compressed` | **Left wrist** (eye-in-hand) | Close-up of the left gripper and the fabric it is working on. |
| `/camera_right_wrist/color/image_raw/compressed` | **Right wrist** (eye-in-hand) | Close-up of the right gripper and the fabric it is working on. |

Suggested LeRobot names if you convert this MCAP: `camera_high`, `camera_low`, `camera_left_wrist`, `camera_right_wrist`.

A typical episode has:

- ~2,000 frames per camera
- ~6,500–8,000 `joint_cmd` / `joint_states` messages
- ~40k messages in total
- ~60–80 seconds of wall time

### Training fields

- **observation:** `camera_high` + `camera_low` + `camera_left_wrist` + `camera_right_wrist` + `joint_states`
- **action:** `joint_cmd` (identical names / order / units as `joint_states`)
- **language:** `Fold the T-shirt on the table.`

## License

Apache License 2.0

## Citation

If you use this sample, please mention **TeleXperience** and the dataset name `FoldingTShirt_DualArxR5a_Samples`.