Robust Localization, Mapping, and Navigation for Quadruped Robots
Dyuman Aditya, Junning Huang, Nico Bohlinger, Piotr Kicki, Krzysztof Walas, Jan Peters, Matteo Luperto, Davide Tateo · 2025
Read Robust Localization, Mapping, and Navigation for Quadruped Robots by Dyuman Aditya, Junning Huang, Nico Bohlinger, Piotr Kicki, Krzysztof Walas, Jan Peters, Matteo Luperto, Davide Tateo (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-02272.txt.
Introduction
Dyuman Aditya and co-authors build a navigation stack for low-cost four-legged robots in 2025. Legged robots are now cheap and sturdy, yet real-world use needs solid navigation on cheap sensors. Depth cameras alone must carry the load in their setup.
The stack fuses contact-aided leg odometry, visual-inertial odometry, and depth-stabilized vision. That blend steadies position estimates when single cues fail. Tests in simulation and on two real robot models yield accurate 2D maps, stable self-location, and solo navigation. Ablation tests weigh each part of the stack. Code and videos sit on the project site for others to check.
Catalog record
| CREATOR | Dyuman Aditya, Junning Huang, Nico Bohlinger, Piotr Kicki, Krzysztof Walas, Jan Peters, Matteo Luperto, Davide Tateo |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 39 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-02272
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Robust Localization, Mapping, and Navigation for Quadruped Robots Dyuman Aditya, Junning Huang, Nico Bohlinger, Piotr Kicki, Krzysztof Walas, Jan Peters, Matteo Luperto, Davide Tateo Abstract Quadruped robots are currently a widespread platform for robotics research, thanks to powerful Reinforcement Learning controllers and the availability of cheap and robust commercial platforms. However, to broaden the adoption of the technology in the real world, we require robust navigation stacks relying only on low-cost sensors such as depth cameras. This paper presents a first step towards a robust localization, mapping, and navigation system for low-cost quadruped robots. In pursuit of this objective we combine contact-aided kinematic, visual-inertial odometry, and depth-stabilized vision, enhancing stability and accuracy of the system. Our results in simulation and two different real-world quadruped platforms show that our system can generate an accurate 2D map of the environment, robustly localize itself, and navigate autonomously. Furthermore, we present in-depth ablation studies of the important components of the system and their impact on localization accuracy. Videos, code, and additional experiments can be found on the project website. 1 1 1 https://sites.google.com/view/low-cost-quadruped-slam I INTRODUCTION The widespread availability of robust and reliable robotics platforms, and the major advances in locomotion controllers based on Reinforcement Learning [ 1 , 2 , 3 , 4 , 5 , 6 ] , has enabled the deployment of quadruped robots in challenging environments, both outdoor and indoor [ 7 , 8 ] . Concurrently, there is an increasing interest in quadruped robots from the industry for surveillance, inspection, and search and rescue tasks, owing to their ability to navigate diverse settings—including stairs, cluttered spaces, and uneven terrain. Unfortunately, practical industrial solutions typically rely on expensive robotic platforms equipped with high-precision sensors that enable robust autonomous navigation and mapping. In low-cost platforms, however, these sensors are not available, and most of the localization capabilities must rely on inexpensive depth image cameras or 2D LiDARs, resulting in subpar performance. These issues are further exacerbated when employing policies learned with Reinforcement Learning, which can achieve very high speeds and reactive behavior, at the expense of trajectory smoothness. Such rapid trajectories may disrupt the SLAM pipeline, resulting in loss of localization, broken maps, and, consequently, the inability to navigate autonomously in the environment. These challenges are even more relevant when quadruped robots operate in indoor environments, thus in settings where fast movements change the robot’s perceptions rapidly; in those environments, even the availability of expensive long-range LiDARS has a limited impact on performance, due to the proximity of obstacles to the robot’s location. Fig. 1: Overall system design for robust navigation in low-cost quadruped robots. Our contributions are highlighted in light blue, different colors indicate different types of data. In this paper, we focus on designing a simple and robust localization pipeline that enables low-cost quadruped robots to navigate autonomously in complex environments, thereby empowering these robots to tackle interesting real-world tasks. Our approach targets scenarios where the robot is equipped with a low-cost RGBD camera and an IMU. Due to the type and quality of sensors available for our target set-up, we restrict our localization and navigation pipeline to indoor settings. These environments, while being one of the target environments for quadruped robots, are particularly challenging as fast robot movements can cause abrupt changes in robot perception, thus causing localization and mapping issues. Moreover, we focus on localization, navigation, and mapping using 2D grid maps. Those maps are commonly used for path planning, localization, and path execution since widespread methods for 2D map-based navigation are still de facto the state-of-the-art for autonomous navigation [ 9 ] , and since 2D maps can be combined with machine learning techniques and 3D and semantic maps [ 10 ] . In a 2D mapping setting, an occupancy map is generated using LiDAR scans or from depth data obtained by an RBGD camera [ 11 ] . While setup is widely used in classical wheeled mobile platforms with stable performance, its implementation with quadruped robots is still challenging, and few robust solutions have been proposed yet. Due to rough, uneven, and slippery terrain, robot perception data are subject to jitters and noise, making 2D mapping and localization highly unreliable. Scans oscillate according to the terrain and the robot’s own movement, especially when the movement involves sharp changes in roll and pitch angles. Several works used 3D map representation and vision-based methods for SLAM with quadruped robots. Unfortunately, while 3D maps can provide a more semantically rich representation of the environment, they are still more computationally expensive to obtain and maintain. Moreover, when low-cost sensors are used, 3D maps are heavily affected by noise, making them impractical for navigation. To enhance system autonomous navigation performance, we exploit joint encoder data to compute leg odometry [ 12 ] . This integration requires reliable contact information, typically obtained from contact sensors. However, these sensors are not always available or sufficiently accurate on legged platforms. Recognizing that contact sensors are not always available or sufficiently accurate on legged robotic platforms, we also demonstrate how to estimate foot contacts using only joint torque measurements. Additionally, we incorporate a scan stabilization technique to mitigate the impact of rapid, dynamic movements on the SLAM pipeline, ensuring more robust and accurate localization. The key contributions of this paper are i) a detailed description of the localization system, illustrating how it is possible improve localization and navigation performance integrating standard pipelines with specific problem-dependent solutions; our system involves a) an estimation of legged odometry and b) contact estimation, c) scan stabilization, and d) integration of leg-visual-inertial odometry to improve visual odometry. Then, we provide (ii) an extensive ablation study of the impact of each component of the localization system in simulation in several downstream tasks such as mapping, navigation, and exploration. Finally, we present (iii) an experimental real-world evaluation including a robot with an actuated spine. II Related works Quadruped robots, similarly to other autonomous mobile robots, need localization, mapping, and autonomous navigation capabilities [ 13 ] . While many different systems with such capabilities have been presented in literature, most of them rely on expensive robotic platforms, such as the Anymal and the Spot robots, mounting high-end, expensive long-distance sensors, thanks to their large form factor, allowing heavier payloads. An example of such systems can be found in [ 14 ] for the Anymal and in [ 15 ] for the Spot robot. Indeed, in [ 15 ] , the authors make a step towards low-end platforms by reducing the total amount of sensors used by the system. However, their setup is still relatively expensive compared with the setup considered in this paper. On top of this, we consider much more reactive and fast neural-based policies. A similar setup is used in [ 13 ] , where a long-distance 3D LiDAR and RGBD cameras are used for 3D SLAM in outdoor environments, while the recent work of Li et al. [ 16 ] presents a framework to perform 3D SLAM integrating features extracted from a long-distance LiDAR. In [ 17 ] , the authors describe a framework designed to solve the task of autonomous exploration for object search, where a Spot robot is equipped with a Velodyne Puck LiDAR and an Azure Kinect