CAPO-LeggedRobotOdometry 🦾

Language / 语言： English | 中文

CAPO-LeggedRobotOdometry is a pure proprioceptive odometry library for legged robots, implemented with a portable C++ estimator core that depends only on IMU and joint motor data.

The core estimation logic is implemented in FusionEstimator/fusion_estimator.h. The file fusion_estimator_node.cpp provides a ROS 2 wrapper around this estimator, while the Matlab/ folder contains examples for MATLAB + C++ mixed compilation and offline evaluation.

For side-by-side comparison, Matlab/Comparison/invariant-ekf/ provides a MATLAB mixed-compilation workflow for invariant-ekf, making it easier to compare this repository against a representative open-source legged odometry baseline.

In 3D closed-loop trials (a 200 m horizontal and 15 m vertical loop), Astrall point-foot robot A achieves 0.1638 m horizontal error and 0.219 m vertical error; for wheel-legged robot B, the corresponding errors are 0.2264 m and 0.199 m.

Welcome to join the wechat group

---

📄 Paper

Contact-Anchored Proprioceptive Odometry for Quadruped Robots (arXiv:2602.17393)

• Paper: https://arxiv.org/abs/2602.17393

If you use this repository in research, please consider citing the paper.

---

📦 Data Sharing (Go2-EDU ROS bags)

To help readers quickly validate the pipeline, we provide Go2-EDU trial datasets, including real-world videos, derived CSV files, and the corresponding ROS bag topics/messages required by this node, enabling fast reproduction and sanity checks.

• Download (GitHub Releases): https://github.com/ShineMinxing/CAPO-LeggedRobotOdometry/releases/tag/DataForTest
• Recommended assets in that release include GO2Flat, GO2Stairs, MPXY150Z10, MWXY150Z10, robot_flat_1_compress.zip, and robot_stairs_1_compress.zip

Note: the IMU on this Go2-EDU platform exhibits noticeable yaw drift, so the odometry accuracy is generally worse than the results reported for Astrall robots A and B in the paper.

---

✨ Key Features

Category	Description
Biped / Quadruped / Wheel-Legged Unified	Online contact-set switching; stance legs are detected automatically, supporting fast transitions between standing and walking.
IMU + Joint-Motor Only	The estimator core works with only IMU and joint motor measurements, without requiring cameras or LiDAR.
MATLAB / C++ Mixed Compilation	The Matlab/ folder provides MATLAB + MEX examples for calling the same C++ core, and Matlab/Comparison/invariant-ekf/ includes a comparable mixed-compilation setup for invariant-ekf.
Full 3D & Planar 2D	Publishes both 6DoF odometry (SMX/Odom) and a gravity-flattened 2D odometry (SMX/Odom_2D).
Portable Pure C++ Core	The estimator core is isolated in FusionEstimator/, making it easier to reuse outside ROS2.
Runtime Tuning	Key parameters can be adjusted through config.yaml, and platform-dependent thresholds can be tuned for different robots.

---

🏗️ Related Repositories

Scope	Repository	Summary
Low-level / Drivers	https://github.com/ShineMinxing/Ros2Go2Base	DDS bridge, Unitree SDK2 control, pointcloud→LaserScan, TF utilities
Odometry	CAPO-LeggedRobotOdometry (this repo)	Pure proprioceptive fusion, publishes SMX/Odom / SMX/Odom_2D
SLAM / Mapping	https://github.com/ShineMinxing/Ros2SLAM	Integrations for Cartographer 3D, KISS-ICP, FAST-LIO2, Point-LIO
Voice / LLM	https://github.com/ShineMinxing/Ros2Chat	Offline ASR + OpenAI Chat + TTS
Vision	https://github.com/ShineMinxing/Ros2ImageProcess	Camera pipelines, spot / face / drone detection
Gimbal Tracking	https://github.com/ShineMinxing/Ros2AmovG1	Amov G1 gimbal control and tracking
Tools	https://github.com/ShineMinxing/Ros2Tools	Bluetooth IMU, joystick mapping, gimbal loop, data logging

⚠️ Clone as needed. If you only need state estimation, this repository is sufficient. For mapping, it is natural to pair it with Ros2SLAM and Ros2Go2Base.

---

📂 Repository Layout

CAPO-LeggedRobotOdometry/
├── CMakeLists.txt
├── package.xml
├── config.yaml
├── fusion_estimator_node.cpp        # ROS2 wrapper around the C++ estimator core
├── FusionEstimator/                 # portable pure C++ estimator core
│   ├── Estimators/
│   ├── fusion_estimator.h           # main estimator entry
│   ├── LowlevelState.h
│   ├── SensorBase.cpp
│   ├── SensorBase.h
│   ├── Sensor_IMU.cpp
│   ├── Sensor_IMU.h
│   ├── Sensor_Legs.cpp
│   ├── Sensor_Legs.h
│   └── Readme.md
├── Matlab/                          # MATLAB + MEX examples for the same C++ core
│   ├── build_mex.m
│   ├── fusion_estimator.m
│   ├── fusion_estimator_mex.cpp
│   ├── Comparison/
│   │   └── invariant-ekf/           # MATLAB mixed-compilation workflow for invariant-ekf
│   └── ...                          # optional test datasets are published via GitHub Releases
├── Plotjuggler.xml
└── Readme.md

---

🧩 Architecture Notes

This repository is intentionally split into three layers:

1. Estimator core

The actual odometry algorithm is implemented in FusionEstimator/fusion_estimator.h and the accompanying files under FusionEstimator/.
This is the main pure C++ proprioceptive estimator core, designed to work with only IMU and joint motor data.

2. ROS2 wrapper

fusion_estimator_node.cpp wraps the estimator core into a ROS 2 node, handling:

• ROS2 topic subscriptions
• parameter loading
• message conversion
• odometry publication

3. MATLAB mixed-compilation support

The Matlab/ folder provides examples for compiling and calling the same C++ estimator core from MATLAB via MEX.
In addition, Matlab/Comparison/invariant-ekf/ provides a comparable MATLAB mixed-compilation workflow for invariant-ekf, which is useful for offline benchmarking and side-by-side evaluation.

---

⚙️ Configuration (config.yaml)

The table below lists the main parameters used by the ROS2 node. See config.yaml for the full file and comments.

Parameter	Typical Value	Meaning
sub_imu_topic	SMX/Go2IMU	IMU topic
sub_joint_topic	SMX/Go2Joint	joint state topic
sub_mode_topic	SMX/SportCmd	reset / mode topic
pub_odom_topic	SMX/Odom	6DoF odometry output
pub_odom2d_topic	SMX/Odom_2D	planar odometry output
odom_frame	odom	odometry frame
base_frame	base_link	body frame
base_frame_2d	base_link_2D	planar body frame
imu_data_enable	true	enable IMU input
leg_pos_enable	true	enable leg-position kinematics
leg_vel_enable	true	enable leg-velocity kinematics
leg_ori_enable	false	enable kinematics-based yaw correction
leg_ori_init_weight	0.001	initial weight of orientation correction
leg_ori_time_wight	1000.0	time-related growth weight for orientation correction
leg_velcke_enable	false	enable IKVel-CKF / CAPO-CKE
foot_force_threshold	20.0	contact threshold
min_stair_height	0.08	minimum stair-height hypothesis used by the estimator

Notes: * The exact values are platform-dependent. * Some fields in config.yaml may be experimental, legacy, or intended for specific branches / workflows. * If you port only the pure C++ core, you usually need to keep only the estimator-relevant fields and can drop ROS2-specific naming.

---

⚙️ Build & Run

Clone the repository

mkdir -p ~/ros2_ws/src
cd ~/ros2_ws/src
git clone --depth 1 https://github.com/ShineMinxing/CAPO-LeggedRobotOdometry.git
cd ~/ros2_ws
colcon build --packages-select fusion_estimator
source install/setup.bash
ros2 run fusion_estimator fusion_estimator_node

📑 ROS 2 Interfaces

/fusion_estimator_node (rclcpp)
├─ Publishes
│   • SMX/Odom         nav_msgs/Odometry (odom → base_link)
│   • SMX/Odom_2D      nav_msgs/Odometry (odom → base_link_2D)
├─ Subscribes
│   • SMX/Go2IMU       sensor_msgs/Imu
│   • SMX/Go2Joint     std_msgs/Float64MultiArray
│       layout:
│         data[0..11]   = 12 joint positions q
│         data[12..23]  = 12 joint velocities dq
│         data[24..27]  = 12 joint torque tau or contact-related values
│   • SMX/SportCmd     std_msgs/Float64MultiArray
│       reset example:
│         data[0] == 25140000  → estimator reset
└─ TF is typically published by another node / utility if needed

---

🧠 Algorithm Overview (High Level)

1. Contact detection
   Detect stance legs from force / threshold logic.

2. Forward kinematics
   Compute foot-end position / velocity in the body frame from joint measurements.

3. Contact anchoring
   Record touchdown footfall points and use them as intermittent world-frame constraints during stance.

4. Height stabilization
   Use support-plane height logic and confidence decay to reduce long-horizon elevation drift.

5. Optional IKVel-CKF (CAPO-CKE)
   Suppress encoder-induced velocity spikes and obtain smoother leg-end velocity estimates.

6. Optional yaw stabilization
   Use multi-contact geometric consistency to reduce IMU yaw drift during prolonged standing.

7. 2D projection
   Publish a planar odometry (SMX/Odom_2D) by flattening roll / pitch while preserving yaw and planar motion.

For full derivations, please refer to the paper.

---

🔧 Reusing FusionEstimator/ Outside ROS2

If you do not need ROS2, the reusable part is mainly under FusionEstimator/.

A typical migration path is:

1. keep the estimator core in FusionEstimator/
2. replace ROS2 message subscriptions with your own sensor interface
3. prepare IMU, joint position / velocity, and contact / force inputs
4. call the estimator core from your own loop
5. export odometry to your target middleware / runtime

This design is useful for:

• ROS1 migration
• embedded deployment
• offline replay tools
• MATLAB / simulation analysis
• custom robotics middleware

---

🧪 MATLAB / MEX (Optional)

The Matlab/ folder provides examples for compiling the estimator core into a MATLAB-callable MEX module.

Typical workflow:

cd Matlab
build_mex
fusion_estimator

Files in this folder:

• build_mex.m — build script for MEX compilation
• fusion_estimator_mex.cpp — MEX bridge wrapping the C++ estimator core
• fusion_estimator.m — MATLAB-side usage / validation example
• sample CSV files and compressed test datasets — published via GitHub Releases: https://github.com/ShineMinxing/CAPO-LeggedRobotOdometry/releases/tag/DataForTest

For comparison with another representative legged-odometry implementation, Matlab/Comparison/invariant-ekf/ also provides a MATLAB mixed-compilation workflow for invariant-ekf.

---

📈 Visualization

A Plotjuggler.xml file is included in the repository for convenient visualization and debugging with PlotJuggler.

You can use it to inspect:

• odometry outputs
• IMU-related signals
• joint / force-related channels
• estimator behavior during walking, standing, or reset events

---

🎥 Videos

Topic	Link
Odometry-only mapping (morphology switching)
Indoor walking (0.5%–1% error)
Stair climbing (height error < 5 cm)
Outdoor walking (380 m, 3.3% error)
Voice interaction + navigation
Face tracking + laser spot tracking
AR glasses head-following
YOLO drone tracking
Gimbal + fixed camera collaboration
Multiple SLAM integrations

---

📨 Contact

Email	Affiliation
sunminxing20@mails.ucas.ac.cn	Institute of Optics and Electronics, CAS

This repository is under active development. Issues and PRs are welcome.

---

CAPO-LeggedRobotOdometry 🦾 (中文)

CAPO-LeggedRobotOdometry 是一个面向腿式机器人的纯本体感知里程计库，核心是**只依赖 IMU 和关节电机数