To move over, around, or through obstacles in the world, robots and animals need to employ a repertoire of dynamic and dexterous behaviors. Since the world is ever-changing, these behaviors must be synthesized on-the-fly and adapted to diverse environmental conditions. At present, animals deftly outperform autonomous robots in this regard. We seek tools that will enable the performance of dynamic legged robots to surpass that of their animal counterparts.
In this talk, we discuss advances in modeling and control of dynamic legged locomotion. Unlike some areas of robotics and biomechanics, models for most dynamic legged behaviors have poor predictive power. In particular, rigid-body models of legged locomotion yield predictions that vary discontinuously when multiple limbs contact terrain. By introducing compliance in limbs, we show that model predictions vary smoothly with respect to initial conditions (including states, parameters, and inputs). Smooth model predictions are amenable to scalable algorithms for estimation, optimization, and learning; we briefly discuss our current efforts and future plans in these directions. We conclude that compliance in limbs perform morphological computations that can simplify modeling and control of dynamic legged locomotion.