README.md

---
license: mit
task_categories:
- robotics
language:
- en
tags:
- robot-learning
- industrial-assembly
- learning-from-demonstration
- autoregressive model
- mixture of experts
pretty_name: ATG-MoE Pressure-Reducing Valve Assembly Training Set
size_categories:
- n<1K
---

# ATG-MoE Pressure-Reducing Valve Assembly Training Set

[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
[![Paper: arXiv](https://img.shields.io/badge/arXiv-2603.19029-b31b1b.svg)](https://arxiv.org/abs/2603.19029)
[![Project Page](https://img.shields.io/badge/Project-Page-blue)](https://hwh23.github.io/ATG-MoE)
[![GitHub](https://img.shields.io/badge/Code-GitHub-181717?logo=github)](https://github.com/hwh23/ATG-MoE)

This repository provides the **training set** used in our paper: **"ATG-MoE: Autoregressive trajectory generation with mixture-of-experts for assembly skill learning"**. 

The dataset is specifically designed for the **pressure-reducing valve assembly task**, featuring multi-skill robotic learning capabilities.

> [!IMPORTANT]
> This release only contains the **training set**.  
> Evaluation requires a Unity-based simulation environment. The evaluation scripts and instructions will be released in our [GitHub repository](https://github.com/hwh23/ATG-MoE).

---

## 📌 Dataset Overview

The dataset supports **multi-skill imitation learning** in industrial assembly. It provides high-quality demonstrations mapping multi-view observations to precise skill trajectories.

### 🛠️ Included Skills
The dataset covers **8 key skills** required for the assembly process:
* Sleeve Placement
* Large Spring Insertion
* Rod Placement
* Rod Seating
* Nut Seating
* Spring Insertion
* Plug Seating
* Body Seating

### 📊 Dataset Scale
* **Total Training Episodes:** 768
* **Skills:** 8 distinct assembly tasks
* **Data per Skill:** 96 training episodes

---

## 📂 Data Format

The dataset follows the **RLBench** organization style. Each `episode` folder contains a complete demonstration trajectory.

### Directory Structure
```text
episode0/
├── front_depth/                # Depth maps from front camera
├── front_rgb/                  # RGB images from front camera
├── left_shoulder_depth/        # Depth maps from left shoulder camera
├── left_shoulder_rgb/          # RGB images from left shoulder camera
├── overhead_depth/             # Depth maps from overhead camera
├── overhead_rgb/               # RGB images from overhead camera
├── right_shoulder_depth/       # Depth maps from right shoulder camera
├── right_shoulder_rgb/         # RGB images from right shoulder camera
├── proprioception/             # Robot joint states and end-effector pose
├── camera_matrix.json          # Intrinsic and extrinsic parameters
├── kfs.json                    # Keyframe indices for trajectory
├── lang_emb.pkl                # Language instruction embeddings
├── variation_descriptions.json # Natural language descriptions
└── variation_number.pkl        # Variation ID for the task
```

---

## 📝 Citation
If you find this dataset or our work helpful, please cite:

```bibtex
@misc{huang2026atgmoeautoregressivetrajectorygeneration,
  title={ATG-MoE: Autoregressive trajectory generation with mixture-of-experts for assembly skill learning},
  author={Weihang Huang and Chaoran Zhang and Xiaoxin Deng and Hao Zhou and Zhaobo Xu and Shubo Cui and Long Zeng},
  year={2026},
  eprint={2603.19029},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2603.19029},
}