To maximize the cost-effectiveness of neural network (NN) accelerators, architects are actively developing single-chip accelerators which can execute many NNs simultaneously. However, previous approaches fail to achieve full performance potential by exploiting only spatial or temporal resource sharing (SS or TS). They also do not consider memory management that can significantly affect performance. This limitation leads to the dire need for a new multi-NN accelerator taking both opportunities with careful memory management. But, it is extremely challenging to design an ideal spatio-temporal sharing accelerator because it requires (1) an algorithm that determines the degree of SS/TS in large exploration spaces, (2) a new STS-enabled accelerator devised with diverse design points, and (3) carefully-designed memory management that minimizes resource contention during numerous data transfers upon reconfiguration. To this end, we propose STfusion, a fast and flexible multi-NN execution architecture. First, STfusion partitions an accelerator into multiple smaller TS-enabled accelerators. Second, STfusion dynamically fuses small accelerators to adjust the accelerator sizes. Third, STfusion manages on-chip buffer in a page-granularity for stall-free data transfers. Lastly, STfusion provides an algorithm that determines the degree of SS/TS to achieve high throughput while satisfying QoS goals. Our evaluation shows that STfusion significantly outperforms state-of-the-art multi-NN accelerators.

Recently developments in the field of telecommuni-cations and network technology has been formed new paradigm shift in defense weapon system. The military also developing the Tactical Information Communication Network (TICN) system to meet the changing aspects of these wars. To equip its function from the battlefield, TICN system have to be received real-time information from the battlefield troop, and the key technologies that support these features is MANET (Mobile Ad-hoc Network) routing protocol. In this paper, basis on the tactical environment of small combat units, by studying on the MANET routing protocol that can be transmitted in real time & reliable information, suggest efficient routing protocol in tactical MANET that can be applicable to advanced digital battlefield.

In this paper, a new biomimetic control method for a quadruped walking robot is proposed. The method is derived by the observation of the gravity load receptor and stimulus- reaction mechanism of quadrupeds' locomotion, and the study of the stances on walking and energy efficiency. Though the controller is simple, it provides a useful framework for controlling a quadruped walking robot. In particular, by introducing a new rhythmic pattern generator the heavy computational burden to be paid on solving kinematics is relieved. The effectiveness of the proposed method is validated via a dynamic simulation and experimental works in a quadruped walking robot, called AiDIN(Artiflcial Digitigrade for Natural Environment).

One of the traditional problems in the walking and climbing robot moving in the 3D environment is how to negotiate the boundary of two plain surfaces such as corners, which may be convex or concave. In this paper a practical gait planning algorithm in the transition region of the boundary is proposed in terms of a geometrical view. The trajectory of the body is derived from the geometrical analysis of the relationship between the robot and the environment. And the position of each foot is determined by using parameters associated with the hip and the ankle of the robot. In each case of concave or convex boundaries, the trajectory that the robot moves along is determined in advance and the foot positions of the robot associated with the trajectory are computed, accordingly. The usefulness of the proposed method is confirmed through simulations and demonstrations with a walking and climbing robot.

When a robot navigates in the unstructured environment, it frequently meets various objects capable of being represented as a simple geometrical primitives such as planes, spheres, cylinders, and their combinations. In this paper, we propose a method for the perception of the environment with a low cost two degree-of-freedom IR (infrared) scanner, which can be used for navigation of the quadruped walking robot. Our work is mainly focused on the development of an algorithm for real-time modeling of the environment and the shape reconstruction of the environment. Though the proposed algorithm is very simple, it is useful to determine the geometrical features of the object as well as provide sufficient information for navigation of the robot. Using the proposed method, a method for gait control of the quadruped walking robot can be generated. The validity of the proposed algorithms are demonstrated through experiments using a quadruped walking robot called "MRWALLSPECT III(Multifunctional Robot for WALL inSPECTion version 3)".

This paper describes an integrated approach for the design of a quadruped robot, called "MRWALLSPECT III (multifunctional robot for wall inspection version 3)", which walks in planes as well as climb walls with suction pads. Since the robot is expected to be able to move from wall to wall to plane including negotiating convex or concave corners, a comprehensive study is needed to develop the robot by performing intuitive and geometrical analysis, critical aspects of design such as joint ranges, design of ankles, and location of actuators are discussed. Environmental geometries are classified into several groups, and criteria for design are proposed for each case. The proposed criteria are applied to the actual design of MRWALLSPECT III designed to carry an ultrasonic NDT tool for inspection of the large surface of industrial utilities. By implementing the proposed idea on the robot its effectiveness is experimentally confirmed.