Graduate School of Science and Engineering Information and Computer Science

Research Field (A):
Software Agents: Robot Controllers, Car Drivers, and Entities in Multi-agent Systems

Research Topic (A1): Emergence of Social Behavior in Multi-agent Systems

Research Topic (A2): Evolving Optimal Swimming Gaits of Fish Robot

Background

The research on the fish robot (Fishbot, Figure 3) was initially inspired by the recent discovery in live science that the undulatory locomotion gaits of leaving creatures that move in (water, or air) fluids exhibit a rather narrow (optimal) range of the Strouhal number. Initially, we attempted to incorporate such a domain-specific knowledge in order to prune the potentially huge search space of parameters of undulation in order to achieve a computationally efficient optimization of the swimming speed and energy consumption of the bot.

Algorithmic Paradigm

The mentioned above computational efficiency of optimization is important, because, as we intended to bridge the reality gap (caused by the challenge of simulating realistically the fluid vortices), we implemented an evolutionary (e.g. involving many trials) optimization via genetic algorithms in a real time, directly on the physical bot, rather than its software model. Moreover, such an approach also requires appropriate solutions to the typical challenges in evolving physical robots, such as mechanical robustness, power supply, automated fitness measurement, and automated repositioning of the bot between the trials.

Current Research

a) Morphology: discovering the optimal stiffness, and optimal layout of the caudal fin, introducing a limitedly-underactuated caudal joint (Figure 4), elastic actuators, etc.

b) Control of undulations: optimizing the main parameters of the canonical sinusoidal control (Figure 5a), introducing various non-sinusoidal control patterns – such as hysteretic (Figure 5b), exponential (Figure 5c), triangular (Figure 5d), etc.

Research Topic (A3): Self-driving Cars

Background

We commenced our research on self-driving cars back in 2004, when we “drove” an off-the-shelf scale RC racing car via a PC-running driving agent. The latter used an overhead video camera to obtain its sensory information, and an RC transmitter to convey the throttle-, brake- and steering commands to the car (Figure 6a). As a result of the proposed anticipatory modeling of the car, we realized a stable (non-oscillatory) control even at the maximum speed of the car (2 m/s, scaled up to about 170 km/h) despite the delayed (about 300 ms) sensory feedback from the camera. Also, we achieved both a precise and consistent control by implementing a real-time (sampling interval: 33 ms) video tracking of the car. Moreover, we parameterized (formalized) the driving style (i.e., the racing line, and the speed breakdown along the line) of the car and employed an evolutionary computation (genetic algorithms, and genetic programming) to evolve the optimal driving style – i.e., the style that results in a fastest lap time around various O-, 8-, and S-shaped test tracks. Since then, we have conducted a research on self-driving cars on various platforms and environments, such as scale RC cars in an augmented reality environment (Figure 6b), the open source car racing simulator (TORCS, Figure 6c), and – recently – a fully-fledged drive simulator (Figure 7). The latter allowed us to extend our research beyond the control of the car, and to include both (i) the psychological aspects pertinent to human drivers in self-driving cars, and (ii) the aids that could support the drivers in emergency situations.

Current Research

(a) Control: The research on control is dedicated to providing both (i) stable (non-oscillatory) and (ii) comfortable control of the self-driving cars. Regarding the former, we are researching models that incorporate the variable feedback delays in the lateral (i.e., steering), and longitudinal (braking and accelerating) control during the lane keeping (LK), lane changing (LC), obstacle avoidance (OA), and adaptive cruise control (ACC) of the car. To achieve a comfortable control (which, in turn, would address the challenge of motion sickness in self-driving cars) we consider minimizing the lateral- and longitudinal jerks of the car. We are interested in investigating various control models of LK, LC, OA, and ACC, and optimizing these models (e.g., via machine learning) from the viewpoints of both stability and comfortability of control.

(b) Human factors: It is essential that the human driver in Levels 3 and 4 of automated driving could quickly and safely take over the control of the car when requested. Such a take-over request (TOR) would be a result of either an unrecognized traffic situation by the auto pilot, or a contradictory (or lack of sufficient) sensory information provided to the latter. The psychological phenomena that could compromise the performance of the driver before- or during TOR include cognitive underload, highway hypnosis, automation surprise, and automation startle. We are researching on modalities (text, voice, vibration), and types of human-machine interfaces (HMI) – such as transparent HMI, semantic HMI, personal HMI, and external HMI – that could alleviate the effects of the above-mentioned psychological phenomena (Figure 8).

(c) Advanced driving aids: During TOR, as elaborated above, the quick response of the drivers is crucial. In most cases, the reaction of the drivers during TOR would involve either an emergency steering, or (and) an emergency braking. In the latter case, it is important to reduce the driver reaction time, consequently – the reaction distance, and the corresponding overall stopping distance of the car. To accomplish this, the existing emergency brake assist (EBA) systems sense the onset of the movement of the brake pedal (or, the increase of the hydraulic pressure in the brake system), and, if a sudden movement (or a sudden increase of pressure) is being detected, automatically apply the maximum amount of braking force (Figure 9a). The objective of our research is to further reduce the driver reaction time by detecting early the movement of the driver’s right foot (from its resting position on the floor) towards the brake pedal, classifying the time series of such a movement (as either a normal- or an emergency braking) via machine learning, and activating the brakes well before the driver ever reaches to the brake pedal (Figure 9b). For the detection of the movement of driver’s foot we are investigating sensory subsystems based on video feed, pressure sensor, etc.

Research Field (B):
Intelligent Control of Future Wireless Communication Systems

In future wireless networks, referred to as beyond 5G, a vast and diverse array of information and communication devices—including not only smartphones but also Internet-of-Things (IoT) devices, vehicles, drones, and satellites—will connect to the network as “users.” To accommodate this explosive growth in user devices and to ensure efficient and seamless communication, we are conducting research and development on foundational technologies for future communication networks, leveraging a variety of theoretical frameworks, including stochastic modeling, control theory, and machine learning.

Research Topic (B1): Intelligent Adaptive Network Control for the Spatiotemporal Dynamics of Users

At outdoor concerts and sports events, the phenomenon of "hot spots"—where user density is spatially concentrated—is commonly observed. Moreover, the locations of these hot spots change over time due to user mobility. To address such spatiotemporal variations in user distribution, the application of new network systems and control technologies, such as millimeter and terahertz wave communications, beamforming, massive multiple-input-multiple-output (MIMO), reflecting intelligent surface (RIS), and aerial base stations, is expected to provide users with a comfortable communication experience. However, how to control these diverse technologies in real time remains a significant challenge. For example, the use of unmanned aerial vehicles (UAVs) as aerial base stations has recently attracted attention not only for addressing hot spots but also for deployment during disasters or large-scale network failures. However, the optimal deployment of UAVs in three-dimensional space—while considering the communication quality of all users and interference among UAVs—remains a challenging problem. In response, we are conducting research on the deployment and trajectory control of aerial base stations using various techniques, including mathematical optimization, distributed control theory, game theory, and applied probability, to enable adaptation to the spatiotemporal variations of users.

Research Topic (B2): Performance Evaluation of Next-Generation Networks Based on Probabilistic Modeling

The quality of wireless communication that we usually enjoy in everyday life is affected by a wide range of random factors. These include, for example, inter-cell interference, the mobility of transmitters and receivers, and random signal loss caused by fading and shadowing effects in the wireless channel. Due to the complexity of such wireless environments, comprehensively analyzing their behavior through simulations is often impractical. To address this challenge, performance evaluation methods based on stochastic geometry—a probabilistic modeling approach—have attracted increasing attention in recent years. In this framework, the positions and movements of wireless base stations and user terminals are modeled as spatial stochastic processes (e.g., point processes). This approach enables the theoretical analysis of various performance metrics, such as data rate and the distribution of the signal-to-interference-plus-noise ratio (SINR).

As an example of next-generation wireless communication networks, we have focused on vehicle-to-vehicle (V2V) communication, conducting a theoretical performance evaluation based on stochastic geometry and proposing control methods derived from analytical results. For instance, a cellular vehicle-to-everything (C-V2X) communication network can be modeled as a probabilistic model in which vehicles move along randomly distributed roads, and base stations are distributed according to a spatial point process, as illustrated below. This modeling framework enables a theoretical comparison between cellular-relay-based communication via base stations and direct vehicle-to-vehicle communication. In addition to C-V2X, we are also conducting theoretical performance evaluations of a wide range of emerging network technologies, including millimeter-wave (mmWave) communication networks, non-orthogonal multiple access (NOMA) networks, reconfigurable intelligent surfaces (RIS) networks, and low Earth orbit (LEO) satellite networks.

Research Topic (B3): Data-Driven Network Control Based on Machine Learning Techniques

Recent communication networks, which are used daily by a vast number of users, generate a wide variety of data. On the user side, this includes web access logs and video streaming records, while on the network side, system logs (e.g., syslog) and traffic data can be collected from network devices (e.g., servers, routers, and switches). By analyzing such network-related data using statistical techniques such as machine learning, it is expected that more intelligent operation and control of communication networks will become possible. We are conducting research on automated network monitoring technologies, such as failure prediction and service impact estimation, using log data generated by network devices. We are also developing techniques for inferring user communication behaviors based on network traffic logs, e.g., DNS query logs.