Noise Analysis and Hierarchical Adaptive Body State Estimator For Biped Robot Walking With ESVC Foot

Boyang Chen, Xizhe Zang, Chao Song, Yue Zhang, Xuehe Zhang, Jie ZhaoarXiv 2025 · DOI · arXiv

Abstract

Bipedal robots with curved feet, such as the Ellipse-based Segmental Varying Curvature (ESVC) design, offer improved energy efficiency but introduce significant challenges for state estimation. The rolling contact geometry results in time-varying process noise that standard Extended Kalman Filters (EKF) fail to handle effectively. This paper proposes a hierarchical adaptive state estimator that utilizes a noise-time regression model to dynamically adjust filter parameters. Experimental results on the TT II biped robot demonstrate that this approach achieves faster convergence and higher precision during variable-speed walking compared to conventional methods.

Introduction

Modern bipedal robotics is increasingly adopting bio-inspired curved foot designs to mimic human rollover characteristics. The ESVC foot is a prime example, enhancing efficiency but complicating the kinematics. Because the contact point moves along an elliptical arc, calculating the robot's state involves complex integrals that often require approximation. These approximations, combined with mechanical vibrations, lead to non-constant "process noise" that fluctuates rhythmically with the gait cycle. Standard estimators assuming constant noise often suffer from instability or lag during speed transitions.

Hierarchical Estimation Architecture

To address these challenges, we developed a hierarchical state estimator that separates raw data processing from the adaptive filtering stage.

The TT II Biped Robot platform and the proposed Hierarchical State Estimator architecture.

The architecture consists of three primary layers:

  1. Pre-Estimation Layer: This layer fuses high-frequency data from the IMU, joint sensors, and inclinometers. A specialized two-step velocity fusion process is employed to filter out mechanical noise before the data reaches the main estimator.
  2. Adaptive Post-Estimation Layer: An EKF serves as the core, but its covariance matrices ($Q$ and $R$) are not static. Instead, they are updated in real-time based on a pre-calculated noise model and the current residuals.
  3. Acceleration Fusion: The system integrates theoretical acceleration from a Linear Inverted Pendulum (HLIP) model with actual IMU readings, using a Gaussian kernel to weight the "trust" between the model and the sensors based on the robot's stability.

Noise-Time Regression Model

A key innovation of this work is the treatment of noise as a deterministic, time-varying signal rather than a random constant. By analyzing experimental data from the TT II robot, we identified that process noise follows a predictable pattern during each stride.

Rhythmic noise variance and regression models for joint and IMU sensors.

We utilized weighted cubic polynomial regression to model this variance. As shown in the plots, the variance of sensor data exhibits clear rhythmic fluctuations. By incorporating this regression model into the EKF, the estimator can "anticipate" periods of high noise (e.g., during foot strike or rapid rolling) and adjust its gain accordingly.

Experimental Results

The system was validated using the TT II Biped Robot in tasks involving "marking time" and variable-speed lateral walking.

  • Convergence Speed: During transitions from 0.1 m/s to 0.2 m/s, the proposed estimator stabilized within 4 steps, outperforming the standard EKF which required 6 steps.
  • Tracking Accuracy: Phase portraits of the Center of Mass (CoM) showed that the hierarchical adaptive method tracked ground truth much more closely than baseline filters, particularly during the high-velocity phases of the gait.

Conclusion

The hierarchical adaptive state estimator effectively mitigates the challenges posed by complex foot geometries in bipedal walking. By modeling the rhythmic nature of process noise and employing a multi-stage fusion strategy, the system provides a robust foundation for high-level robot control. Future research will focus on extending this model to handle unstructured terrain and optimizing computational efficiency for real-time deployment on more agile platforms.

Summary updated

BibTeX
@misc{chen2025noise,
  author = {Boyang Chen and Xizhe Zang and Chao Song and Yue Zhang and Xuehe Zhang and Jie Zhao},
  title = {{Noise Analysis and Hierarchical Adaptive Body State Estimator For Biped Robot Walking With ESVC Foot}},
  year = {2025},
  eprint = {2506.08578},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.08578},
}
Figure