Anisotropic compliance of robot legs improves recovery from swing-phase collisions.

Uneven terrain in natural environments challenges legged locomotion by inducing instability and causing limb collisions. During the swing phase, the limb releases from the ground and arcs forward to target a secure next foothold. In natural environments leg-obstacle collisions may occur during the swing phase which can result in instability, and may require contact sensing and trajectory re-planning if a collision occurs. However, collision detection and response often requires computationally- and temporally-expensive control strategies. Inspired by low stiffness limbs that can pass past obstacles in small insects and running birds, we investigated a passive method for overcoming swing-collisions. We implemented virtual compliance control in a robot leg that allowed us to systematically vary the limb stiffness and ultimately its response to collisions with obstacles in the environment. In addition to applying a standard positional control during swing motion, we developed two virtual compliance methods: (1) an isotropic compliance for which perturbations in theanddirections generated the same stiffness response, and (2) a vertical anisotropic compliance in which a decrease of the upwardvertical limb stiffness enabled the leg to move upwards more freely. The virtual compliance methods slightly increased variability along the limb's planned pathway, but the anisotropic compliance control improved the successful negotiation of step obstacles by over 70% compared to isotropic compliance and positional control methods. We confirmed these findings in simulation and using a self-propelling bipedal robot walking along a linear rail over bumpy terrain. While the importance of limb compliance for stance interactions have been known, our results highlight how limb compliance in the swing-phase can enhance walking performance in naturalistic environments.