Model-Based Proprioceptive State Estimation for Spring-Mass Running

Autonomous applications of legged platforms will inevitably require accurate state estimation both for feedback control as well as mapping and planning. Even though kinematic models and low-bandwidth visual localization may be sufficient for fully-actuated, statically stable legged robots, they are inadequate for dynamically dexterous, underactuated platforms where second order dynamics are dominant, noise levels are high and sensory limitations are more severe. In this paper, we introduce a model based state estimation method for dynamic running behaviors with a simple spring-mass runner. By using an approximate analytic solution to the dynamics of the model within an Extended Kalman filter framework, the estimation accuracy of our model remains accurate even at low sampling frequencies. We also propose two new event-based sensory modalities that further improve estimation performance in cases where even the internal kinematics of a robot cannot be fully observed, such as when flexible materials are used for limb designs. We present comparative simulation results to establish that our method outperforms traditional approaches which rely on constant acceleration motion models and that it eliminates the need for an extensive and unrealistic sensor suite.