Generalized state estimation for diverse humanoid motions

Develop a state estimator for humanoid robots that generalizes across diverse, contact-rich motions—including getting up from the ground during motion tracking and recovery behaviors during diffusion-policy control—in order to mitigate state-estimation drift and prevent failures during sim-to-real deployment without relying on motion-specific assumptions about end-effector contacts.

Background

The paper deploys policies on a Unitree G1 humanoid with a 500 Hz state estimator using a generalized momentum observer combined with a Kalman filter, and in some extreme cases incorporates LiDAR–inertial odometry or omits position-dependent observations. Despite these measures, the authors report occasional failures due to state estimation drift, especially when end-effector contact assumptions are violated.

They explicitly identify developing a state estimator that can robustly handle the full variety of contact-rich skills (e.g., get-up motions and diffusion-policy recovery modes) as a remaining challenge, and leave this for future work. Achieving such generalization would reduce sim-to-real failures across the broad motion repertoire demonstrated by their framework.