Humanoid Robots Just Got a 96% Faster Way to Stay Upright on Moving Ground

Researchers Falak Mandali, Zijian He, and Yan Gu from the University of Massachusetts Lowell have introduced a new state estimation method for humanoid robots operating on non-inertial ground, such as moving platforms, ships, or earthquake-affected terrain. The approach uses an invariant extended Kalman filter (InEKF) that relies solely on onboard proprioceptive sensors—specifically IMUs mounted on the robot's feet. By exploiting kinematic constraints at the stance foot, the filter estimates the robot's base position and velocity relative to the moving ground frame without any external measurements or ground motion data.

Experiments on Agility Robotics' Digit humanoid robot demonstrate dramatic improvements over existing InEKFs. On a swaying and pitching ground, the new filter achieved a 96% speedup in convergence rate and an 80% reduction in position estimate errors. In walking tests on a uni-axially rotating ground, the average estimation error remained below 9 cm even when starting with an initial error of up to 1 meter. The observability analysis confirms that relative base position and velocity are observable under the non-inertial ground frame, paving the way for more robust humanoid locomotion in dynamic environments like ships, moving vehicles, or earthquake zones.