MUSE: A Real-Time Multi-Sensor State Estimator for Quadruped Robots

TL;DR

MUSE addresses the challenge of accurate, real-time state estimation for legged robots operating on difficult terrain by integrating proprioceptive and exteroceptive sensing into a five-component pipeline: exteroceptive odometry, attitude observation, slip detection, leg odometry, and sensor fusion. The estimator employs a cascaded attitude framework with a nonlinear observer and an eXogenous Kalman Filter, and fuses IMU and joint data with exteroceptive odometry to produce high-frequency, low-drift pose and velocity estimates. Key contributions include a slip-detection module embedded in the estimator, real-time operation enabling closed-loop control, and extensive validation on both Aliengo (online) and FSC (ANYmal B300) benchmarks, with competitive accuracy and up to 400 Hz operation. The work demonstrates the practical impact of multi-sensor fusion for robust quadruped locomotion and provides an open-source implementation to accelerate further research and deployment.

Abstract

This paper introduces an innovative state estimator, MUSE (MUlti-sensor State Estimator), designed to enhance state estimation's accuracy and real-time performance in quadruped robot navigation. The proposed state estimator builds upon our previous work presented in [1]. It integrates data from a range of onboard sensors, including IMUs, encoders, cameras, and LiDARs, to deliver a comprehensive and reliable estimation of the robot's pose and motion, even in slippery scenarios. We tested MUSE on a Unitree Aliengo robot, successfully closing the locomotion control loop in difficult scenarios, including slippery and uneven terrain. Benchmarking against Pronto [2] and VILENS [3] showed 67.6% and 26.7% reductions in translational errors, respectively. Additionally, MUSE outperformed DLIO [4], a LiDAR-inertial odometry system in rotational errors and frequency, while the proprioceptive version of MUSE (P-MUSE) outperformed TSIF [5], with a 45.9% reduction in absolute trajectory error (ATE).

Figures (6)

• Figure 1: Robot Reference Frames: the navigation frame $\mathcal{N}$, the body frame $\mathcal{B}$, the IMU sensor frame $\mathcal{I}$, the camera frame $\mathcal{C}$, and the LiDAR frame $\mathcal{L}$ for ANYmal.
• Figure 2: MUSE utilizes two exteroceptive sensors (ES): Camera and LiDAR, and three proprioceptive sensors (PS): IMU, encoders, force/torque sensors. It comprises five main components: camera odometry or LiDAR odometry, an attitude observer (AO), slip detection (SD), leg odometry (LO), and a sensor fusion algorithm (SF). The AO includes a nonlinear observer (NLO) and an eXogenous Kalman Filter (XKF). The SD and LO include joint state (JS), robot kinematics/dynamics (KD), ground reaction forces (GRF), and leg odometry models. The SF utilizes a Kalman filter to estimate odometry.
• Figure 3: During the closed-loop experiment, Aliengo walked up and down the stairs, then on rocks and slippery terrain, repeating these tasks three times.
• Figure 4: Aliengo on uneven terrain: Comparison of position and orientation estimations between the GT and MUSE, MUSE without the SD module (MUSE with no SD), Proprioceptive MUSE (P-MUSE), and P-MUSE without the SD module (P-MUSE with no SD). The grey shaded areas indicate that the robot is walking on rocks, while the red ones indicate when the robot is walking on the slippery patch. The position plot (left) shows that the drift is higher when SD is not active.
• Figure 5: Aliengo on uneven terrain: Ground Truth (GT) vs. Linear Velocity estimated by MUSE during the closed-loop experiment, zoom into the time interval [185-230] s.