We explore a model-based approach to reinforcement learning where partially or totally unknown dynamics are learned and explicit planning is performed. We learn dynamics with neural networks, and plan behaviors with differential dynamic programming (DDP). In order to handle complicated dynamics, such as manipulating liquids (pouring), we consider temporally decomposed dynamics. We start from our recent work [1] where we used locally weighted regression (LWR) to model dynamics. The major contribution of this paper is making use of deep learning in the form of neural networks with stochastic DDP, and showing the advantages of neural networks over LWR. For this purpose, we extend neural networks for: (1) modeling prediction error and output noise, (2) computing an output probability distribution for a given input distribution, and (3) computing gradients of output expectation with respect to an input. Since neural networks have nonlinear activation functions, these extensions were not easy. We provide an analytic solution for these extensions using some simplifying assumptions. We verified this method in pouring simulation experiments. The learning performance with neural networks was better than that of LWR. The amount of spilled materials was reduced. We also present early results of robot experiments using a PR2. Accompanying video: https://youtu.be/aM3hE1J5W98.

The Human-Robot Interaction gets increasingly closer. In consequence, human safety has become a key issue for the success of the symbiosis between humans and robots. When the minimum distance between a human and a robot is too short, it can be naturally considered that the probability of a collision increases. Therefore, we consider that the robot should increase the distance to the human when the human is getting closer. We propose Withdrawal strategy as a method that aims to increase the distance by moving the end-effector not only away from the human but also to a parking position that can be previously assessed to be safer. To withdraw the end-effector, we use a virtual force model consisting of two virtual forces: a repelling force exerted by the human and an attractive force exerted by the parking position. We carry out experiments using a human-sized humanoid robot and five human subjects, and report the task completion time to evaluate the efficiency of the robot when performing a simple task.

Continuous action sets are used in many reinforcement learning (RL) applications in robot control since the control input is continuous. However, discrete action sets also have the advantages of ease of implementation and compatibility with some sophisticated RL methods, such as the Dyna [1]. However, one of the problem is the absence of general principles on designing a discrete action set for robot control in higher dimensional input space. In this paper, we propose to construct a discrete action set given a set of basis functions (BFs). We designed the action set so that the size of the set is proportional to the number of the BFs. This method can exploit the function approximator's nature, that is, in practical RL applications, the number of BFs does not increase exponentially with the dimension of the state space (e.g. [2]). Thus, the size of the proposed action set does not increase exponentially with the dimension of the input space. We apply an RL with the proposed action set to a robot navigation task and a crawling and a jumping tasks. The simulation results demonstrate that the proposed action set has the advantages of improved learning speed, and better ability to acquire performance, compared to a conventional discrete action set.

This paper presents a new continuous action space for reinforcement learning (RL) with the wire-fitting. The wire-fitting has a desirable feature to be used with action value function based RL algorithms. However, the wire-fitting becomes unstable caused by changing the parameters of actions. Furthermore, the acquired behavior highly depend on the initial values of the parameters. The proposed action space is expanded from the DCOB, proposed by Yamaguchi et al., where the discrete action set is generated from given basis functions. Based on the DCOB, we apply some constraints to the parameters in order to obtain stability. Furthermore, we also describe a proper way to initialize the parameters. The simulation results demonstrate that the proposed method outperforms the wire-fitting. On the other hand, the resulting performance of the proposed method is the same as, or inferior to the DCOB. This paper also discuss about this result.