Progress towards a humanoid robot that learns to stand

Most of us are able to stand and likely learned to do so in the first year of our lives. While it may look easy, it is actually a difficult problem in sensory-motor control that involves multiple senses. A robot can be programmed to stand but its behavior will be more robust if it goes through an incremental learning process. Such an approach is characteristic of the emerging field of developmental or epigenetic robotics [1][2]. A mobile robot that can sense the state of its body and its environment is well suited to reinforcement learning or learning through trial and error [3]. Other types of learning based on training or imitation [4] are possible but a trial and error approach allows the robot to learn on its own and gives insight into human learning. Applying reinforcement learning to humanoid robots is challenging because their bodies have many degrees-of-freedom which leads to an enormous amout of trial and error. This paper proposes three simplifications to facilitate the application of reinforcement learning. The idea of applying reinforcement learning to robots has been explored by others [5][6]. The use of reinforcement learning in continuous spaces was addressed in [7]. Some related applications of reinforcement learning such as walking [8] and robot soccer [9] have been reported.