Synthesis of stylized walking controllers for planar bipeds

We present a new method for generating controllers for physics-based simulations of planar bipedal walking, targeted towards producing autonomous character behaviors for computer animation. A common criticism of physics-based character animation is the lack of personality and style in the final motion. We develop controllers that mimic a desired style as much as possible, while still subject to the laws of physics and to the realities of maintaining balance. The resulting simulated character can interact with and respond to its environment, unlike the original kinematically-specified desired motion. Walking is a very challenging control task because of its dynamically unstable nature. Continuous balance control is required to prevent the character from falling over or reaching poses from which it cannot recover. The complexity of a walking controller is compounded by high dimensional continuous state and action spaces. Our controller implements a nearest-neighbor control policy with respect to a set of nodes embedded in the state space, and is parameterized by the locations of these nodes and their associated actions. We use greedy search to find an optimized controller given an intuitive objective function. Importantly, several shaping techniques are used to ensure that local minima in the optimization space define good, meaningful solutions. We demonstrate the ability to generate stable walks of various styles, walks for bipeds of varying dimensions, and walks that are robust to unobserved terrain variations.