HR2-KILO: A High-Rate, Robust, Kinematic-Inertial-LiDAR Odometry for Humanoid Robots
Jixin Gao, Fusheng Zha, Lianzhao Zhang, Wei Guo, Pengfei Wang, Lining Sun, Mantian Li
AI summary
Problem
Humanoid locomotion is inherently unstable and non-smooth, demanding high-frequency state estimation that existing methods struggle to provide without relying on redundant contact sensors or suffering from high drift.
Approach
The system tightly couples LiDAR, IMU, and joint encoder data within a manifold error-state Kalman filter, using a pointwise update strategy and online contact detection based on IMU fluctuations and foot clearance.
Key results
- Achieves 6–15 kHz odometry output rate
- Reduces drift and improves accuracy over state-of-the-art methods
- Enables contact detection without physical sensors or dynamic models
- Validated across multiple humanoid platforms and diverse real-world scenarios
Why it matters
Provides a lightweight, high-frequency localization solution that meets real-time control demands, accelerating the practical deployment of dynamic humanoid robots.
Abstract
In this letter, we present a high-rate and robust multi-sensor fusion framework for state estimation of hu- manoid robots, named HR2-KILO. To handle the inherently high-dynamic characteristics of humanoid robots, the proposed framework tightly couples the measurements from the joint encoder, inertial sensor, and LiDAR. We estimate states within the error-state Kalman filter, incorporating the pointwise update strategy, IMU measurement model, and multiple leg kinematic information. Moreover, acceleration fluctuations, foot positions, and the history map are utilized for online contact detection without any contact sensors. The overall system fully utilizes the available multi-source information, making it compact and easy to deploy. Extensive experiments are conducted both on the public dataset and in the real world, including different humanoid robots and diverse scenarios. The results demonstrate that HR2-KILO achieves extremely high rate output and lower drift compared to state-of-the-art LiDAR-inertial(-kinematic) methods. To contribute to the community, the source code and the multi-sensor humanoid dataset are released.