PRVR: Learning Posture Recovery for a Legged Spherical Robot Using Vector Reward

Abstract

Legged spherical robots feature both walking and rolling modes and they are useful in explorations, but often end up in various body-inverted postures after rolling and accidental falls. Posture recovery in such cases where all legs are in the air and lose leg-ground contacts is an open and challenging problem. We propose a posture recovery learning method using vector reward called PRVR. A traditional scalar reward cannot evaluate each element of a vector action. Here, a vector reward method is proposed to evaluate each element of a vector action, such that the action of each joint closely conforms to the expected behavior. To reduce learning complexity, a rotation symmetry-based training and deployment method is proposed. Body-inverted postures can be divided into several symmetric classes, and then through training on one class of body-inverted postures, the robot can recover from various body-inverted postures. Simulations and experiments on a six-legged spherical robot are used to verify the effectiveness.