README.md

---
dataset_name: MoSim Dataset
language:
- en
license: mit
tags:
- reinforcement-learning
- world-model
- dynamics-simulation
- motion
- state-action
- npz
task_categories:
- reinforcement-learning
pretty_name: MoSim Dataset
size_categories:
- 1K<n<10K
---

# 🗂️ MoSim Dataset

Official release of the dataset from the paper:  
**[Neural Motion Simulator: Pushing the Limit of World Models in Reinforcement Learning](https://arxiv.org/pdf/2504.07095)**  

This dataset contains sequential **state-action trajectories** for training and evaluating **MoSim (Neural Motion Simulator)** world models.  
All trajectories are collected **from random policies** in classical control and locomotion environments.

---

## 📦 Dataset Overview

- **Format**: `.npz` (NumPy compressed arrays)  
- **Contents**:  
  - `*_random.npz`: training episodes  
  - `*_random_test.npz`: test episodes  
- **Episode length**: 1000 steps per episode  

---

## 📊 Data Structure

Each `.npz` file contains:

| Key        | Shape                               | Description                        |
|------------|-------------------------------------|------------------------------------|
| `states`   | *(num_episodes, 1000, state_dim)*    | State at each timestep             |
| `actions`  | *(num_episodes, 1000, action_dim)*   | Action applied at each timestep    |

**State composition**:

1. **Joint DOF (articulated body)**  
   - Joint angles *(radians)*  
   - Joint angular velocities *(rad/s)*  
2. **Root DOF (global free body)**  
   - Root position *(x, y, z)*  
   - Root linear velocity *(vx, vy, vz)*  
3. **Root Orientation & Rotation**  
   - Root rotation quaternion *(qx, qy, qz, qw)*  
   - Root angular velocity *(wx, wy, wz)*  

> ⚡ For manipulation control tasks like `Panda`, only joint angles and velocities are provided.  
> ⚡ For locomotion tasks like `Humanoid` or `Go2`, full root DOF and velocities are included.

---