Studies on autonomous mobile robot navigation frequently employ crowd models within simulations to study robot movement in crowded environments. While numerous crowd models exist, each possesses distinct strengths and weaknesses, making no single model universally applicable. Consequently, selecting an appropriate model for a specific scenario presents a significant challenge. To address this issue, our research focuses on a method to classify simulation environments into categories and select a suitable crowd model based on this classification. As a key component of this approach, this paper proposes a methodology to evaluate the suitability of crowd models. This involves comparing pedestrian movement trajectories extracted from real-world datasets with those generated by crowd simulations conducted in equivalent virtual environments. Specifically, we represent the variations within sets of movement trajectories as probability distributions. The similarity between the distribution derived from the real-world dataset and the distribution generated by the simulation serves as the evaluation metric. Furthermore, we present evaluation experiments using real-world datasets and a crowd simulator to validate the effectiveness of the proposed evaluation approach.

Crowd models are utilized in various fields, including research on autonomous mobile robots, where they serve as an experimental environment for selecting and evaluating appropriate navigation methods. It is important to note that while there are numerous crowd models available, each has its own strengths and weaknesses, and no single model is universally applicable to all scenarios. Therefore, selecting the appropriate model for a crowded environment can be a challenging task that requires careful consideration and evaluation. Against this background, we have studied methods to classify scenarios for simulations into multiple classes and select appropriate crowd models based on the classification results. In this paper, a method is proposed for evaluating various crowd models in this context. The paper presents a method for comparing pedestrian movement trajectory data from a dataset with the results of a crowd simulation in an equivalent environment. The comparison is carried out using the visual shape of the movement trajectories and Dynamic Time Warping. The effectiveness of the method is validated by evaluating the impact of the crowd model and its parameter values on the accuracy of simulating a particular scenario.

We study software automatic tuning. Automatic tuning tools using iterative one-dimensional search estimate hyperparameters of machine learning programs. Iterative one-dimensional search searches the parameter space consisting of possible values of the parameters to be tuned by repeatedly measuring and evaluating the target program. Since it takes time to train a machine learning program, estimating the optimal hyperparameters is time-consuming. Therefore, we propose a method to reduce the time required for automatic tuning by parallelization of iterative one-dimensional search. For parallelization, we use multiple job execution on a supercomputer that can utilize multiple GPUs, which is effective for machine learning. In this method, each job measures different hyperparameters. The next search point is determined by referring to the data obtained from each job. The target program is a pedestrian path prediction application. This program predicts future routes and arrival points based on past pedestrian trajectory data. The program is intended to be used in a variety of locations, and the locations and movement patterns will vary depending on the dataset used for training. We hypothesized that the estimation results of one dataset could be used for automatic tuning of another dataset, thereby reducing the time required for automatic tuning. Experimental results confirm that the parallelized iterative one-dimensional search reduces the estimation time from 89.5 hours to 4 hours compared to the sequential search. We also show that the iterative one-dimensional search efficiently investigates the point at which the performance index improves. Moreover, the hyperparameters estimated for one data set are used as the initial point for the search and automatic tuning for another data set. Compared to the results of automatic tuning with the currently used hyperparameters as the initial values, both the number of executions and execution time were reduced.

In recent years, to achieve safe and efficient navigation for autonomous mobile robots, several methods have been proposed to switch between multiple policies (action decision methods), including deep reinforcement learning, depending on the situation. We have also proposed methods to introduce new policy-switching criteria and to add new policies to avoid freezing conditions. Compared to existing methods, we have shown that the method improves safety metric values (e.g., collision rate) even in narrow corridors; however, there are still challenges in achieving sufficient performance because safety and efficiency metric values fluctuate depending on the environment. In this paper, we propose an adaptive navigation method that uses sensing results to classify robot deployment environments into several groups and adaptively changes the policy-switching algorithm according to the environment. Specifically, we use collision risk and congestion level for the environment classification and associate the environment classes with appropriate control parameter values (i.e., parameter tuning) to achieve adaptive navigation. Furthermore, we verify the effectiveness of the proposed method by conducting simulation experiments.

In recent years, many navigation methods using deep reinforcement learning for autonomous mobile robots have been proposed to apply to various dynamic environments. However, since the learning results depend on the simulation environment, it may be inappropriate to apply them directly to the real environments from the standpoint of safety and efficiency. To solve the problem, in this paper, we propose a multi-policy switching method that enables safe and efficient navigation for autonomous mobile robots that share space with humans in various real-world environments, such as dense and crowded situations. Specifically, the method switches the navigation rule among four policies including deep reinforcement learning-based policy according to the size of the target robot’s movable space (i.e., unoccupied area around the robot). We verify the effectiveness of the proposed method by conducting navigation experiments with computer simulation. The results show that the proposed method improves collision avoidance rate in a narrow space where the robot tends to halt or oscillate by existing methods.

Robot simulators are generally used in the research and development of autonomous mobile robots in human-robot coexistence environments. These simulators require to implement moving pedestrians as dynamic obstacles, and often use various crowd models developed in the field of crowd simulation research. However, existing crowd simulations make assumptions that all virtual agents comprising the crowd are controlled in the same manner and that each virtual agent has full knowledge of the surrounding environment. The validity of these assumptions is not clear in robot navigation simulators. In this paper, we analyze existing crowd simulations for use in a simulator for autonomous mobile robot navigation. We also conduct simulation experiments using a simple crowd model to evaluate the impact of environmental differences on the performance of the crowd simulations.

In this paper, we propose a path planning method using reinforcement learning for addressing the navigational issues of an autonomous mobile robot that shares space with a person and travels to its destination safely and efficiently. Here, we generate a path that does not interfere with the pedestrian’s path by incorporating the relative positions of the robot and the pedestrian, as well as the moving tendencies of the target pedestrian (velocity and direction) into the learning state. Simulation results showed that, compared to existing methods (e.g., artificial potential methods), the proposed method does not block the pedestrian’s path and takes a safe route to avoid collisions, although the duration to reach the goal was longer.

Until now, we have been studying a method for predicting the future trajectory of a pedestrian using a machine learning algorithm for the purpose of improving the tracking accuracy of a human-following mobile robot. Here, an open dataset was used to generate the predictor during the training phase, and the future trajectory was predicted by using just one sensor on the robot during the prediction phase. However, the specific method of constructing the training data was not considered, and there was a problem that sufficient accuracy was not obtained in the case of selecting inadequate datasets. To solve the problem, in this paper, we propose a method of extracting similar sets of data from an open dataset by expressing the features of the data in the target environment as a probability distribution and evaluating the divergence between the source distribution and the target distribution. Specifically, we express the features of a set of data as a multidimensional Gaussian distribution and compare the similarity between the distributions using the Kullback-Leibler divergence. In order to verify the effectiveness of the proposed method, we conduct an evaluation experiment using an LSTM-based prediction model as a machine learning algorithm. The results show that we can express the similarity of the movement tendency of pedestrians in the dataset by the Kullback-Leibler divergence based on the target dataset and that the prediction accuracy increases as the value is smaller.

Recent years have seen an increase in the use of data acquired by sensors and wearable devices. However, depending on the type of sensor or wearable device, the data may be irregular with missing data, outliers, and different units of measurement. The use of these data as direct input into a machine-learning model would not produce the correct results. Therefore, analysts would be required to pre-process the data before data analysis to obtain accurate results. In particular, sensor data may contain more outliers and missing data because of network congestion and the limited life of sensor batteries than data acquired by other means. To efficiently perform such preprocessing, we previously proposed APREP-S (automatic preprocessing of sensor data) using Bayesian inference based on programming by example. APREP-S defines one model for each imputation method, as the workflow selects models based on the features of the imputation area. Therefore, this APREP-S model must be regenerated when data with a different periodicity are used. In other words, depending on whether the data are affected by the weekday or weekend, weather conditions, seasons, etc., the imputation model would have to be generated to consider these features. In this study, we enhanced the method for selecting the optimal imputation model in APREP-S, allowing multiple models to be defined for each input method. We evaluated APREP-S, which uses two types of data, by the mean squared error of these data: 1) human activity data as short-term periodic data, and 2) temperature and humidity data as long-term periodic data. As a result, we concluded that APREP-S is an efficient imputation method.

It has recently become possible to analyze integrated data obtained from sensors or wearable devices in the field of Information Technology (IT). Well-known examples include behavioral patterns of customers in a shop, the autonomous motion of robots, and fault prediction. It is known that the pre-processing of data - before it is input into the analysis model - is essential to achieve accurate results. However, this task requires 80% of the resources in the analysis process. As an improvement to the pre-processing of outliers and missing data, we had previously proposed “APREP-S” based on Bayesian inference with Programming by Example. In this paper, we expand APREP-S and define the timing of model generation for increased interactivity between analysts and APREP-S. We calculate the accuracy of the new APREP-S and compared it to that of existing methods. Based on our results, we conclude that APREP-S is the most effective imputation method for time-series data.

In this paper, we propose a method robustly predicting the destination of a pedestrian heading toward a robot in order to provide suitable voice guidance to him/her by communication robots installed at the reception desks of public facilities. For this purpose, we measure a pedestrian trajectory with a laser range scanner attached to the robot, and predict the destination among more than three branches by cascading multiple predictor models for two branches pre-trained by a machine learning algorithm. In order to verify the effectiveness of the proposed method, we conduct experiments using a dataset of tracking pedestrians at a shopping mall, and data observed in the real environment. The result shows that our method can predict three branch destinations with an accuracy of about 80%.

This paper proposes a method to predict the future trajectory of a pedestrian as sequence data by using massive trajectory records collected by various sensor devices. We aim to use the method for safely and efficient path planning of autonomous mobile robots in a human-robot coexisting environment. For the prediction, we use a sequence-to-sequence model, which is frequently used in the field of natural language processing and enables to treat long-term sequence data. In order to verify the effectiveness of the proposed method, we conduct experiments using a dataset of tracking pedestrians at a shopping mall. The result shows that our method can predict sequences sufficiently by converting the trajectory data to adequate sequence data.

Recently, lifestyle-related diseases, which are caused by accumulated disordered lifestyle, have been a considerable problem. Though someone might be able to change such lifestyle by becoming aware of the style problems, it is not easy to be always conscious of health in everyday life and also to visualize their lifestyle individually. Currently, we can use various sensor devices to solve these problems. Improving of sensing technologies makes it possible to collect various kinds of vital information daily and easily. From these backgrounds, we have studied on a Quality of Life (QOL) visualization system, whose target is to improve our lifestyle on a daily basis by using vital information collected from wearable devices. In this paper, we propose a QOL indicator called SRV (Stress and Relaxed Value) as one of the functions of the QOL visualization system. That is an estimation model on stress and relaxed states by using the instantaneous pulse rate. We also verify the effectiveness of the proposed model by conducting evaluation experiments using a pulse sensor.

We have studied a dialog control method for interface robots by using location data of persons measured by a Laser Range Scanner as a human-robot interaction technology. In the method, we measured the distance between a sensor and a person with the sensor placed at human waist height as a time series data and estimated the position coordinates of the person at a time as a probability distribution. This paper extends the scheme and proposes a method estimating person attributes in addition to the location data by monitoring the movement of legs while the person is walking. As for person attributes, we focus on the age and classify persons as the elderly and the young. At that time, we construct a prediction model of age groups based on machine learning mechanisms. In this paper, we use seven feature values, these are the step length, the step width, the velocity of leg 1, the velocity of leg 2, the velocity of body, the acceleration of leg 1 and the acceleration of leg 2 for the model. By conducting experiments, we verify that classification accuracy improves particularly using acceleration and standard deviation of the data.

This paper proposes a remote navigation system for a simple tele-presence robot. In general, robot functions of tele-presence robots are limited, and visual angle of the camera on robots is narrow, so they are unsuitable to use navigation services. The proposed system overcomes the difficulty by adopting system architecture in which robots are controlled remotely via a server located in the Internet, by using well-designed virtual reality techniques for robot control and by implementing almost all robot navigation functions into the server. In this paper, we also implemented Remote Open Campus System (ROCS) as a prototype system, and by using that, we confirmed the effectiveness and usefulness of the proposed system.

In recent years, there are great attentions to realworld data services where the Internet services, ubiquitous computing and robot services are integrated. In order to spread such services widely, standardized platforms and wide variety of applications are required. In this paper, we propose a Remote Open Campus System (ROCS) which provides a campus tour service by controlling a robot at a campus through the Internet. ROCS maps a real world campus environment to a virtual world, that is, reflecting real views captured by a robot to a virtual campus created by itself. For that, ROCS functions are created by using RSNP (Robot Service Network Protocol), which is a protocol specification for robot services.

Internet and cloud-based robot services and their platforms are becoming attractive. Many related works have been also done, such as DAvinCi, ROS on Android, and RoboEarth. However, when robot services are provided via computer networks, very reliable services are required to deal with short/long-term disconnection between services and robots due to wireless LAN disconnection, robot service problems, system error on robots, and so on. In this paper, we adopt Robot Service Network Protocol (RSNP) as a robot service platform in order to integrate robot services with Internet services. In addition, we propose a method and architecture to realize reliable cloud-based robot services for RSNP in a communication view point.

With the rapid advance of the integration of internet and robot areas, various service platforms are proposed that assume cloud environments. For robot services, as the numbers of devices and their types increase, a mechanism that enables the entry of developers in various areas is required. However, existing platforms are insufficient to the problem. In this paper, we developed a distributed service framework that realizes the coordination of various devices, robots, and service functions based on RSNP (Robot Service Network Protocol), a protocol specification for robot services. Furthermore, we implement a pet-sitting service as a prototype system that uses a proposed framework and verify the effectiveness of the framework.

Recently, many ICT companies as well as researchers are taking an increasing interest in robot services using cloud computing, and various service platforms for them have been proposed. From these backgrounds, we have proposed a distributed service framework using Robot Service Network Protocol (RSNP) to integrate various devices including robots with internet services. The key mechanism of the proposed framework is a robot assignment function, which discovers distributed robot resources and assign the requested tasks by end users to suitable robots. In this paper, we implement a prototype system on the proposed framework, and conduct implementation evaluation using the system.

We, RSi (Robot Service Initiative) organization, have been developing a common network based robot service platform, named RSNP (Robot Service Network Protocol) since 2004. As spreading actual use of RSNP, strong requirements are raised on the push communication in limited conditions such as fewer operators and/or limited resources, and on the robot service integration with various devices supported by the other robot platform, such as RTM (Robot Technology Middleware), particularly. In this paper, we clarified these requirements and solved them by pseudo PUSH communication method, by introducing multimedia/sensor profile and by building RSi/RTM gateway. Moreover, we evaluate the effectiveness of the proposed scheme through the performance experiments. And also these results have been also reflected in RSNP 2.0, the latest specification.

This paper analyzes relationship between video contents and subjective video quality for Internet broadcasting, and applies the results to the rate control methods. In this analysis, we classify video programs into some groups in which a large majority of users feel the same subjective video quality. We investigate the relationship by conducting an experiment with method of paired comparisons, and express that as equations by multiple linear regression analysis.

We propose a streaming video system, which can be used in various system environments. This system has three features for providing real-time video delivery services in the Internet. The first is to use QoS (quality of service) control rules adapting to the system environments. The second is to improve estimation accuracy of the network conditions while a user is watching a stream. The third is to notify QoS degradation factors to the user actively. These features make it possible to provide streaming services with a high level of user satisfaction. We show an experimental system in our laboratory as an implementation example.

As a large number of distributed multimedia systems are deployed on computer networks, quality of service (QoS) for users becomes more important. This paper defines it as application QoS, and proposes the application QoS management system (QMS). It controls the application QoS according to the system environment by using simple measurement-based control methods. QMS consists of three types of modules. These are a notificator for module detecting QoS deterioration, a manager module for deciding the control method according to the application management policies, and a controller module for executing the control. The QMS manages the application QoS by communicating between these modules distributed on the network. Moreover, this paper especially focuses on the function setting QoS management policies to the QMS and proposes the setting method. By a simulation experiment, we confirmed that the system made it possible to negotiate the QoS among many applications and it was able to manage the whole applications according to the policies.

The asynchronous transfer mode (ATM) is a key technology for broadband integrated services digital networks (B-ISDNs), which require high speed transmission. We propose a "self-sizing network operation". This is a traffic engineering, management and operation concept for ATM networks. This concept allows networks to be rapidly operated and flexibly re-dimensioned by the system. We overview the functions of this system. Next, we present the virtual path (VP) bandwidth control function and the network element (NE) interface, which are Step 1 functions in the "self-sizing network operation". Their functions can automatically adjust the VP bandwidth. We show their effectiveness; the VP bandwidth can be reduced by 30% or 40%. To implement the operation system, we considered the actual operation time. This may be restricted for operation control. Then, we evaluated the performance of the Step 1 function. In this evaluation, we measured the basic CMIP performance, and estimated the time necessary for actual operation based on these measurements. A major portion of the total operation time is the time necessary for collecting data. This increases by the cube of the total number of virtual channel handlers (VCHs) with every additional VCH. If we use this system, we must consider its operation cycle and effectiveness. However, fortunately the time necessary for operation is short enough to operate about 50 VCHs in a network which cover all of Japan.