README.md

---
license: gpl-3.0
viewer: false
task_categories:
- robotics
---

# JEPLO Dataset

This dataset accompanies the paper [JEPLO: Joint-Embedding Predictive Learning for LiDAR-Based Legged Locomotion](https://huggingface.co/papers/2609.15770).

**Authors:** Qihao Yuan, Yixuan Qiu, Ziyu Cao, Ming Cao, and Kailai Li.

**Project website:** [JEPLO](https://asig-x.github.io/jeplo_web/)

**Code:** [ASIG-X/JEPLO](https://github.com/ASIG-X/JEPLO)

**Video:** [YouTube](https://youtu.be/HekLtX37ijs?si=uoCrwZoS4Y4LlFco)


This dataset is recorded onboard a Unitree Go2 using a Livox Mid360 LiDAR running a perceptive locomotion policy enabled by JEPLO. The dataset can be used for evaluating legged odometry systems in challenging scenarios using proprioceptive and exteroceptive sensors. 

## Dataset Overview

Publicly released to support the legged robotics community, the dataset comprises **eight sequences** recorded during our experiments: **five indoor sequences** and **three mixed indoor–outdoor sequences**.

The sequences were recorded with a Unitree Go2 quadruped equipped with a Livox Mid360 LiDAR mounted upside down on the robot's head. The dataset provides complete proprioceptive and exteroceptive recordings, including:

- Livox Mid360 LiDAR point clouds.
- Measurements from the LiDAR's built-in IMU.
- Joint encoder measurements.
- 6-DoF ground-truth trajectories.


## Ground Truth

Indoor ground-truth 6-DoF poses are recorded at **100 Hz** in a **10 m × 4 m** test area using **eight Qualisys Miqus M3 motion-capture cameras**.

Ground-truth trajectories are provided for each sequence in the `gt/` directory. Each trajectory is stored in **TUM trajectory format**, with one pose per line:

```text
timestamp tx ty tz qx qy qz qw
```

The ground truth represents the **6-DoF pose of the LiDAR frame**. The fields `tx`, `ty`, and `tz` describe translation, and `qx`, `qy`, `qz`, and `qw` describe orientation as a quaternion.

## Custom ROS 2 Message

Joint encoder measurements are published using the custom ROS 2 message `unitree_msgs/msg/LowStateStamped`. The message definition is available in `unitree_msgs/msg/`.

`LowStateStamped` is a timestamped variant of the standard Unitree Go2 `LowState` message. It contains the robot's low-level state, including joint encoder measurements.

## Sensor Extrinsics

The extrinsics below specify transformations **from the named source frame to the LiDAR frame**. Quaternion values use the order **`(qw, qx, qy, qz)`**, and translations are in **meters**.

### Built-in IMU Frame to LiDAR Frame

```yaml
q_li: [1, 0, 0, 0] # (qw, qx, qy, qz)
t_li: [0.011, 0.02329, -0.04412] # meters
```

### Robot Base Frame to LiDAR Frame

```yaml
q_lr: [0.0, 1.0, 0.0, 0.0] # (qw, qx, qy, qz)
t_lr: [-0.275263, 0.000458, 0.148998] # meters
```

**Quaternion ordering:** The ground-truth trajectory files use `(qx, qy, qz, qw)`, whereas the sensor extrinsics above use `(qw, qx, qy, qz)`.