README.md

---
license: apache-2.0
task_categories:
  - robotics
tags:
  - lerobot
  - robotics
  - cable-insertion
  - manipulation
  - imitation-learning
  - vision-language-action
  - intrinsic
  - ai-for-industry-challenge
  - ur5e
  - sim-to-real
configs:
  - config_name: default
    data_files:
      - split: train
        path: "data/**/*.parquet"
---

# AIC Cable Insertion Dataset

## About the AI for Industry Challenge

This dataset was collected for the [AI for Industry Challenge (AIC)](https://www.intrinsic.ai/events/ai-for-industry-challenge), an open competition by **Intrinsic** (an Alphabet company) for developers and roboticists aimed at solving high-impact problems in robotics and manufacturing.

The challenge task is **cable insertion** — commanding a UR5e robot arm to insert fiber-optic cable plugs (SFP modules and SC connectors) into ports on a configurable task board in simulation (Gazebo). Policies must generalize across randomized board poses, rail positions, and plug/port types.

**Competition Resources**
- **Event Page**: [intrinsic.ai/events/ai-for-industry-challenge](https://www.intrinsic.ai/events/ai-for-industry-challenge)
- **Toolkit Repository**: [github.com/intrinsic-dev/aic](https://github.com/intrinsic-dev/aic)
- **Discussion Forum**: [Open Robotics Discourse](https://discourse.openrobotics.org/c/competitions/ai-for-industry-challenge/)

---

## Dataset Description

This dataset contains teleoperated demonstrations of cable insertion tasks recorded from the AIC Gazebo simulation environment as ROS 2 bag files (.mcap), converted to **LeRobot v2.1** format for training Vision-Language-Action (VLA) policies.

### Key Facts

| Property | Value |
|---|---|
| **Robot** | UR5e (6-DOF) with impedance controller |
| **Simulator** | Gazebo (ROS 2) |
| **Episodes** | 5 |
| **Cameras** | 3 wrist-mounted (left, center, right) |
| **Camera Resolution** | 288×256 (downscaled from 1152×1024 at 0.25×) |
| **FPS** | 20 Hz |
| **Observation State** | 31-dim (TCP pose + velocity + error + joint positions + F/T wrench) |
| **Action Space** | 6-dim Cartesian velocity (linear xyz + angular xyz) |
| **Task Types** | SFP module → NIC port, SC plug → SC port |

### Tasks

Each episode is labeled with a specific language instruction identifying the plug type, target port, and target rail:

| Episode | Task Instruction |
|---|---|
| 0 | Insert the grasped SFP module into sfp_port_0 on the NIC card mounted on nic_rail_0 |
| 1 | Insert the grasped SFP module into sfp_port_0 on the NIC card mounted on nic_rail_2 |
| 2 | Insert the grasped SC plug into sc_port_base on SC port 1 mounted on sc_rail_1 |
| 3 | Insert the grasped SC plug into sc_port_base on SC port 0 mounted on sc_rail_0 |
| 4 | Insert the grasped SFP module into sfp_port_0 on the NIC card mounted on nic_rail_3 |

### Scene Variation

Each trial features different randomization to encourage policy generalization:

| Episode | Board Yaw (°) | Board Height (m) | Cable Type | Other Components Present |
|---|---|---|---|---|
| 0 (Trial 1) | ~25° | 1.140 | sfp_sc_cable | NIC cards on rail 0 & 1, SC mount, SFP mount |
| 1 (Trial 2) | ~45° | 1.200 | sfp_sc_cable | NIC card on rail 2, LC mount, SFP mount |
| 2 (Trial 3) | ~60° | 1.300 | sfp_sc_cable_reversed | SC ports on rail 0 & 1, SFP mount, SC mount, LC mount |
| 3 (Trial 5) | ~15° | 1.110 | sfp_sc_cable_reversed | SC port on rail 0, SFP mounts on both rails |
| 4 (Trial 7) | ~30° | 1.100 | sfp_sc_cable | NIC cards on rail 0 & 3, SC ports on both rails, LC mount, SFP mount |

---

## Data Format and Features

### Observation State (31-dim)

| Index | Feature | Description |
|---|---|---|
| 0–2 | `tcp_pose.position.{x,y,z}` | TCP position in base frame |
| 3–6 | `tcp_pose.orientation.{x,y,z,w}` | TCP orientation (quaternion) |
| 7–9 | `tcp_velocity.linear.{x,y,z}` | TCP linear velocity |
| 10–12 | `tcp_velocity.angular.{x,y,z}` | TCP angular velocity |
| 13–18 | `tcp_error.{x,y,z,rx,ry,rz}` | Tracking error (current vs. reference) |
| 19–24 | `joint_positions.{0–5}` | Joint angles (shoulder_pan → wrist_3) |
| 25–27 | `wrench.force.{x,y,z}` | Wrist force-torque sensor (force) |
| 28–30 | `wrench.torque.{x,y,z}` | Wrist force-torque sensor (torque) |

### Action (6-dim Cartesian velocity)

| Index | Feature | Description |
|---|---|---|
| 0–2 | `linear.{x,y,z}` | Cartesian linear velocity command |
| 3–5 | `angular.{x,y,z}` | Cartesian angular velocity command |

### Camera Views

Three wrist-mounted cameras provide stereo-like coverage of the insertion workspace:

- `observation.images.left_camera` — Left wrist camera (288×256 RGB)
- `observation.images.center_camera` — Center wrist camera (288×256 RGB)
- `observation.images.right_camera` — Right wrist camera (288×256 RGB)

Videos are stored as MP4 files (H.264, 20 fps).

---

## Dataset Structure

```
aic_lerobot_dataset/
├── data/
│   └── chunk-000/
│       ├── episode_000000.parquet
│       ├── episode_000001.parquet
│       ├── episode_000002.parquet
│       ├── episode_000003.parquet
│       └── episode_000004.parquet
├── meta/
│   ├── info.json
│   ├── tasks.jsonl
│   ├── episodes.jsonl
│   ├── episodes_stats.jsonl
│   └── stats.json
└── videos/
    └── chunk-000/
        ├── observation.images.left_camera/
        │   └── episode_00000{0-4}.mp4
        ├── observation.images.center_camera/
        │   └── episode_00000{0-4}.mp4
        └── observation.images.right_camera/
            └── episode_00000{0-4}.mp4
```

---

## Usage

### Loading with LeRobot

```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("shu4dev/aic-cable-insertion")

# Access a frame
sample = dataset[0]
print(sample["observation.state"].shape)   # torch.Size([31])
print(sample["action"].shape)              # torch.Size([6])
```

### Loading with HuggingFace Datasets

```python
from datasets import load_dataset

ds = load_dataset("shu4dev/aic-cable-insertion")
print(ds["train"][0])
```

---

## Data Collection

Demonstrations were collected via **teleoperation** in the AIC Gazebo simulation environment using the LeRobot integration (`lerobot-record`) with keyboard-based Cartesian control. The robot starts each trial with the cable plug already grasped and positioned within a few centimeters of the target port.

Raw ROS 2 bag data (.mcap files, 10–16 GB each) was converted to LeRobot v2.1 format using a custom streaming converter that:

1. Filters to only the 8 needed ROS topics (skipping TF, contacts, scoring)
2. Synchronizes all modalities to the center camera timestamps at 20 Hz
3. Extracts observation state from `/aic_controller/controller_state`, `/joint_states`, and `/fts_broadcaster/wrench`
4. Extracts actions from `/aic_controller/pose_commands` (Cartesian velocity mode)
5. Encodes camera streams as H.264 MP4 via direct ffmpeg pipe

---

## Intended Use

This dataset is intended for:

- Training **imitation learning** policies (ACT, Diffusion Policy, etc.)
- Training **VLA models** (π0, GR00T, OpenVLA, etc.) with language-conditioned cable insertion
- Benchmarking sim-to-sim transfer for contact-rich manipulation
- Research on fine-grained insertion tasks with force feedback

---

## Citation

If you use this dataset, please cite the AI for Industry Challenge:

```
@misc{aic2026,
  title={AI for Industry Challenge Toolkit},
  author={Intrinsic Innovation LLC},
  year={2026},
  url={https://github.com/intrinsic-dev/aic}
}
```

## License

Apache License 2.0