Imitation Learning for Locomotion and Manipulation

Abstract

Decision making in robotics often involves computing an optimal action for a given state, where the space of actions under consideration can potentially be large and state dependent. Many of these decision making problems can be naturally formalized in the multiclass classification framework, where actions are regarded as labels for states. One powerful approach to multiclass classification relies on learning a function that scores each action; action selection is done by returning the action with maximum score. In this work, we focus on two imitation learning problems in particular that arise in robotics. The first problem is footstep prediction for quadruped locomotion, in which the system predicts next footstep locations greedily given the current four-foot configuration of the robot over a terrain height map. The second problem is grasp prediction, in which the system must predict good grasps of complex free-form objects given an approach direction for a robotic hand. We present experimental results of applying a recently developed functional gradient technique for optimizing a structured margin formulation of the corresponding large non-linear multiclass classification problems.