🤖 AI Summary
Legged robots suffer from sensitivity of state estimation to noise covariances and kinematic parameters in dynamic environments, necessitating labor-intensive manual tuning. To address this, we propose a bi-level differentiable optimization framework: the upper level jointly optimizes process/measurement noise covariances and kinematic model parameters, while the lower level executes a full-information closed-loop estimator; gradients are backpropagated through the estimator via the implicit function theorem to enable end-to-end minimization of trajectory-level estimation error. This work is the first to unify noise statistics calibration and kinematic modeling within a differentiable estimation pipeline, ensuring cross-platform generalizability. Experimental validation on quadrupedal and humanoid robots demonstrates significant improvements in state estimation accuracy and uncertainty calibration consistency over manually tuned baselines.
📝 Abstract
Accurate state estimation is critical for legged and aerial robots operating in dynamic, uncertain environments. A key challenge lies in specifying process and measurement noise covariances, which are typically unknown or manually tuned. In this work, we introduce a bi-level optimization framework that jointly calibrates covariance matrices and kinematic parameters in an estimator-in-the-loop manner. The upper level treats noise covariances and model parameters as optimization variables, while the lower level executes a full-information estimator. Differentiating through the estimator allows direct optimization of trajectory-level objectives, resulting in accurate and consistent state estimates. We validate our approach on quadrupedal and humanoid robots, demonstrating significantly improved estimation accuracy and uncertainty calibration compared to hand-tuned baselines. Our method unifies state estimation, sensor, and kinematics calibration into a principled, data-driven framework applicable across diverse robotic platforms.