Postural control in a bipedal robot using sensory reweighting

Postural control is a difficult problem for bipedal robots. Even in robots restricted to the sagittal plane, the system must react to falling forward or backward to stabilize itself during walking, standing, and during initiation and termination of walking. Most robots rely mainly on proprioceptive information such as foot pressure sensors and joint angle sensors for balance. By contrast, humans use a variety of sensory sources, including visual, vestibular, and proprioceptive sources to adapt fluidly to varying conditions. These sensory inputs combine to control posture but are "reweighted" in response to changing conditions such as floor motion, visual scene motion, and degradation in vestibular sensitivity. Based on models of sensory reweighting in humans, we implement a sensory reweighting scheme in a bipedal robot using an adaptive Kalman filter. The adaptive filter uses an online estimate of the noise variance to adjust the Kalman gain depending on time-varying noise conditions. Thus, the robot automatically downweight sensory channels with unreliable data.