In this paper, we propose a complexity-based motion feature learning method and hierarchical spatio-temporal naïve Bayes classifier for human action recognition. As a motion feature learning method, we developed a complexity-based subsequence of time series clustering (C-STSC) method to learn time series codewords from a human motion trajectory. The key to the C-STSC method is to measure the importance of each subsequence in time series data through to use of a complexity measure. Next, time series codewords are learned on the basis of the important subsequences by using a clustering algorithm. Moreover, we also propose a hierarchical spatio-temporal naïve Bayes classifier (HST-NBC) to classify the C-STSC features, where both the codeword-type and its spatio-temporal information is explicitly represented as a composite node in a Bayesian network framework. To validate the proposed method, we present experimental results of the proposed approach with respect to several open datasets.

For a robot to interact with a person effectively, it needs to predict future events that will be caused by the person to occur. By predicting events, a robot can take some preparative actions to reduce the waiting time and greatly improve the interaction. To select the best proactive actions and the best times for those actions, we propose a hybrid temporal influence diagram that can make proactive plans that minimize expected waiting times in a human-robot interaction. To validate our proposed method, we show experimental results for a robotic assistant in a manual assembly task.

For effective human-robot interaction, a robot should be able to make prediction about future circumstance. This enables the robot to generate preparative behaviors to reduce waiting time, thereby greatly improving the quality of the interaction. In this paper, we propose a novel probabilistic temporal prediction method for proactive interaction that is based on a Bayesian network approach. In our proposed method, conditional probabilities of temporal events can be explicitly represented by defining temporal nodes in a Bayesian network. Utilizing these nodes, both temporal and causal information can be simultaneously inferred in a unified framework. An assistant robot can use the temporal Bayesian network to infer the best proactive action and the best time to act so that the waiting time for both the human and the robot is minimized. To validate our proposed method, we present experimental results for case in which a robot assists in a human assembly task.

For a robot to learn complex movement skills, programming by demonstration and/or learning by trial and error is necessary. Measuring the complexity of such movement skills is important to decide the appropriate learning model, and the required size of dataset and additional prior knowledge. To deal with measuring the complexity of movement skills for robots, we propose an information-theoretic complexity measure. By modeling proprioceptive as well as exteroceptive sensory data as a multivariate Gaussian distribution, movement skills can be modeled as a probabilistic model. Next, complexity of the movement skills is measured by using neural complexity. In addition to the original neural complexity measure, endogeneous changes in time of the movement skills are modeled by sampling in time and modeling as individual random variables. To evaluate our proposed complexity measure, several experiments are performed on real robotic movement skills.

In this paper, we propose a proactive assistant robot for human assembly tasks. In order to predict future events of human activities such as requesting parts for assembly, we use the temporal Bayesian network that can infer both causal probability and temporal distribution of an conditional event. Based on the temporal Bayesian network model of an human assembly task, we also show that the proactive assistant robot make human-robot-interaction quickly by temporal prediction of an event.

Abstract — Reinforcement learning (RL) has been used as a learning mechanism for a mobile robot to learn state-action relations without a priori knowledge of working environment. However, most RL methods usually suffer from slow conver-gence to learn optimum state-action sequence. In this paper, it is intended to improve a learning speed by compounding an existing Q-learning method with a shortest path finding algorithm. To integrate the shortest path algorithm with Q-learning method, a stochastic state-transition model is used to store a previous observed state, a previous action and a current state. Whenever a robot reaches a goal, a Stochastic Shortest Path(SSP) will be found from the stochastic state-transition model. State-action pairs on the SSP will be counted as more significant in the action selection. Using this learning method, the learning speed will be boosted when compared with classical RL methods. To show the validity of our proposed learning technology, several simulations and experimental results will be illustrated. I.

The two most important abilities for a robot to survive in a given environment are selecting and learning the most appropriate actions in a given situation. Historically, they have also been the biggest problems in robotics. To help solve this problem, we propose a two-layered action selection mechanism (ASM) which designates an action pattern layer and a reactive behavior plan layer. In the reactive behavior plan layer, a task is selected by comparing behavior motivation values that, in an animal, correspond to external stimuli as well as internal states due to hormones. After a task is selected, its corresponding reactive behavior plan is executed as a set of sequential dynamic behavior motivations (DBMs), each of which is associated with an action pattern. In the action pattern layer, each action pattern can be functionally decomposed into primitive motor actions. Shortest path-based Q-learning (SPQL) is incorporated into both the reactive behavior plan and action pattern layers. In the reactive behavior plan layer, relationships between perceptions and action patterns are learned to satisfy a given motivation, as well as the relative priorities of these relationships. In the action pattern layer, the relations between sensory states and primitive motor actions can be learned. To establish the validity of our proposed ASM, experiments with our real designed robot was illustrated together with simulations.

Reinforcement learning (RL) has been widely used as a learning mechanism of an artificial life system. However, RL usually suffers from slow convergence to the optimum state-action sequence or a sequence of stimulus-response (SR) behaviors, and may not correctly work in non-Markov processes. In this paper, first, to cope with slow-convergence problem, if some state-action pairs are considered as disturbance for optimum sequence, then they are to be eliminated in long-term memory (LTM), where such disturbances are found by a shortest path-finding algorithm. This process is shown to let the system get an enhanced learning speed. Second, to partly solve non-Markov problem, if a stimulus is frequently met in a searching-process, then the stimulus is classified as a sequential percept for a non-Markov hidden state. And thus, a correct behavior for a non-Markov hidden state can be learned as in a Markov environment. To show the validity of our proposed learning technologies, several simulation results are illustrated.