Search

The width of a neural network matters since increasing the width will necessarily increase the model capacity. However, the performance of a network does not improve linearly with the width and soon gets saturated. In this case, we argue that increasing the number of networks (ensemble) can achieve better accuracy-efficiency trade-offs than purely increasing the width. To prove it, one large network is divided into several small ones regarding its parameters and regularization components. Each of these small networks has a fraction of the original one's parameters. We then train these small networks together and make them see various views of the same data to increase their diversity. During this co-training process, networks can also learn from each other. As a result, small networks can achieve better ensemble performance than the large one with few or no extra parameters or FLOPs, \ie, achieving better accuracy-efficiency trade-offs. Small networks can also achieve faster inference speed than the large one by concurrent running. All of the above shows that the number of networks is a new dimension of model scaling. We validate our argument with 8 different neural architectures on common benchmarks through extensive experiments. The code is available at \url{https://github.com/FreeformRobotics/Divide-and-Co-training}., Comment: accepted by T-IP 2022, code is at https://github.com/FreeformRobotics/Divide-and-Co-training

Humans have a remarkable ability to learn continuously from th e environment and inner experience. One of the grand goals of robots is to build an artificial "lifelong learning" agent that can shape a cultivated understanding of the world from the current scene and previous knowledge via an autonomous lifelong development. It is challenging for the robot learning process to retain earlier knowledge when robots encounter new tasks or information. Recent advances in computer vision and deep -learning methods have been impressive due to large-scale data sets, such as ImageNet and COCO. However, robotic vision poses unique new challenges for applying visual algorithms developed from these computer vision data sets because they implicitly assume a fixed set of categories and time -invariant task distributions.

In this paper, a novel soft inflatable arm is proposed for telepresence robots. The new proposed structure of the arm is achieved by a very common and low-cost inflatable material and it is very light, weighing only about 50 g. However, it can realize agile movement by driving six tiny cables installed in the shoulder and elbow joints. The soft inflatable arm can work by pumping air at a very low pressure ( $7.32 \pm 3.45$ kPa) and allows direct and soft human contact without any external force sensors. This paper proposes joint compressed models, joint kinematic models, and kinematic models for the whole inflatable robot arm. These models can be easily applied to multijoint arms. In addition, a new redundancy resolution method is also developed for the inflatable arm, which makes it easier to control resolution and is less complex than other traditional approaches. Numerous experiments have been conducted, including performances of accuracy, repeatability, motion trajectories, human-safe interaction, and remote interaction. Results are satisfactory and validate the expected performance of the proposed robotic arm.

In this paper, we propose a novel design of gaits for space manipulators so as to minimize the dynamic disturbance on a space station. Without modifying their joint configuration, we first propose the use of our previous developed gripping methodology. It combines the concepts of wheels motion in parallel grippers to reduce the swing motions in conventional gaits. For the proposed gaits, three major aspects are considered including 1) linear motion; 2) turning; and 3) exterior transition. Based on the conventional modeling approach for fixed base space manipulators, we extend the method to formulate the dynamic coupling during wheels motion. It is to visualize the manipulator as a prismatic joint with its gripping base moving with wheels motion. Also, we develop an experimental platform for verification. For disturbance analysis, various simulations have been conducted. Results have validated that the proposed gaits generate the lowest disturbance on a space station.

In this paper, we propose an autonomous tree climbing strategy for a novel tree climbing robot that is named Treebot. The proposed algorithm aims to guide Treebot in climbing along an optimal path by the use of minimal sensing resources. Inspired by inchworms, the algorithm reconstructs the shape of a tree simply by the use of tactile sensors. It reveals how the realization of an environment can be achieved with limited tactile information. An efficient nonholonomic motion planning strategy is also proposed to make Treebot climb on an optimal path. This is accomplished by the prediction of the future shape of the tree. The study that is presented in this paper also includes the formulation of Treebot kinematics and an analysis of the workspace of Treebot on different shapes of a tree. Numerous experiments have been conducted to evaluate the proposed autonomous climbing algorithm and to unveil the ability of Treebot.

This paper presents an omnidirectional steer-by-wire system for a four-wheel independent steering vehicle. The system consists of an extended steering interface and a behavior-based steering controller. The extended steering interface provides a novel manipulation way to vehicle driver. The driver can control the vehicle in traditional way or omnidirectionally without any mode switching operation. The reservation of the traditional driving way makes the driver adapt the novel steering interface easily. The steering interface also has a force-feedback function to synchronize the extended steering interface and the orientations of wheels so as to improve vehicle handling. The behavior-based steering controller is designed for controlling the orientations of wheels. It acts as virtual linkage among each wheel to minimize wheel slip resulted by the misalignment of wheels. As there is no trajectory planning needed in the control algorithm, the proposed controller is especially suitable for a nonautonomous vehicle, the driving path of which cannot be predetermined.

In this paper, we present an innovative design of a solder paste inspection system that can be practically integrated into existing solder paste printing machines. Since solder paste inspection systems usually occupy a large vertical space, we designed a mirror box that can redirect the transmission of fringe patterns. In this way, a new parallel-structure solder paste inspection system with a significant reduction in the vertical constraint is developed. We also developed a hybrid weighting algorithm that applies the distance and fringe contrast to acquire the height of solder pastes. Furthermore, we developed an algorithm that generates the 2-D image from the fringe pattern images during the four-step algorithm. It reduces the time required for solder paste inspection compared to traditional approaches that use special lighting systems to create the 2-D image. Based on the results of the height acquisition algorithm, 2-D and 3-D solder paste inspections are performed. Experimental results show that our system can inspect a 20 mm times 20 mm printed circuit board area within 2 s to detect common 2-D and 3-D defects, and the maximum standard deviation for the average height is 3 mum.

In practice, the environments in which mobile robots operate are usually modeled in highly complex forms and, as a result, autonomous navigation and localization can be difficult. The difficulties are exacerbated for practical robots with limited on-board computational resources and complex planning algorithms, since this paradigm of environmental modeling requires enormous computational power. A novel navigation/localization learning methodology is presented to abstract and transfer the human sequential navigational skill to a robotic wheelchair by showing the platform how to respond in different local environments along a demonstrated, static cluttered route using a lookup table representation. This method utilizes limited on-board range sensing information to concisely model local unstructured environments, with respect to the robot, for navigation or localization along the learned route in order to achieve good performance with low on-line computational demand and low-cost hardware requirements. Experimental study demonstrates the feasibility of this method and some interesting characteristics of navigation, localization, and environmental modeling problems. Analysis is also conducted to investigate performance evaluation, advantages of the approach, choices of lookup table inputs and outputs, and potential generalization of this paper

We have developed a single wheel, gyroscopically stabilized robot. This is a novel concept for a mobile robot that provides dynamic stability for rapid locomotion. The robot is a sharp-edged wheel actuated by a spinning flywheel for steering and a drive motor for propulsion. The spinning flywheel acts as a gyroscope to stabilize the robot and it can be steered by tilting. This robot is nonholonomic in nature, underactuated and inherently unstable in the lateral direction. In this paper, we first develop a three-dimensional (3-D) nonlinear dynamic model and investigate the dynamic characteristics of the robot. We conduct simulations and real-time experiments to verify the model. Both simulations and experiments show that the flywheel has a significant stabilizing effect on the robot. Then, we can decouple the longitudinal and lateral motions of the robot by linearization. We propose a linear state feedback to stabilize the robot at different lean angles, so as to control the steering velocity of the robot indirectly, because the robot steers only by leaning itself to a predefined angle. For the task of path following, we design a controller for tracking any desired straight line without falling. In the controller, we first design the linear and steering velocities for driving the robot along the desired straight line by controlling the path curvature. We then apply the linear state feedback to stabilize the robot at the predefined lean angle such that the resulting steering velocity of the robot converges to the given steering velocity. This work is a significant step toward fully autonomous control of such a dynamically stable but statically unstable system.

In this paper, we present a method for modeling human strategies in controlling a light source in a dynamic environment. In order to illustrate the procedure and method, we take a simple example of how to control a light source so as to avoid a shadow and to maintain appropriate illumination conditions on the target area of attention. This work is likely to be valuable in various applications of automatic light control-from surgical rooms and space applications to inspections.

Studies the problem of finite-time output feedback stabilization for the double integrator system. A class of output feedback controllers that can achieve global finite-time stability for the double integrator system are constructed based on a "finite-time separation principle." Furthermore, it is shown that the closed-loop system resulting from our control law can maintain its local finite-time stability regardless of some nonlinear perturbations. Thus, our control law actually applies to a large class of nonlinear second order systems.

This paper studies the problem of modeling and control of multiple cooperative underactuated manipulators handling a rigid object. We reveal holonomic property of such a system by presenting a smooth feedback controller subject to two conditions: 1) there are not fewer active joints than the degrees of freedom of the object; and 2) the Jacobian matrix with respect to passive joints is not singular. This controller is an extension of the PD plus gravity compensation scheme and its asymptotic stability is guaranteed by the LaSalle theorem. Furthermore, we develop a trajectory tracking controller that yields asymptotic convergence of position errors and bounded interaction forces simultaneously. The performance of the proposed controllers has been investigated by simulations on two 6-DOF underactuated manipulators and by experiments on the cooperative underactuated manipulator system developed at CMU.

In this article, we describe and develop methodologies for modeling and transferring human control strategy. This research has potential application in a variety of areas such as the intelligent vehicle highway system, human-machine interfacing, real-time training, space telerobotics, and agile manufacturing. We specifically address the following issues: (1) how to efficiently model human control strategy through learning cascade neural networks, (2) how to select state inputs in order to generate reliable models, (3) how to validate the computed models through an independent, hidden Markov model-based procedure, and (4) how to effectively transfer human control strategy. We have implemented this approach experimentally in the real-time control of a human driving simulator, and are working to transfer these methodologies for the control of an autonomous vehicle and a mobile robot. In providing a framework for abstracting computational models of human skill, we expect to facilitate analysis of human control, the development of human-like intelligent machines, improved human-robot coordination, and the transfer of skill from one human to another.

The authors have developed a unique, single-wheel robot that exploits gyroscopic forces for steering and stability. Experiments with two working models show promise for the concept for high-speed, rough-terrain and amphibious applications.

To successfully interact with and learn from humans in cooperative modes, robots need a mechanism for recognizing, characterizing, and emulating human skills. In particular, it is our interest to develop the mechanism for recognizing and emulating simple human actions, i.e., a simple activity in a manual operation where no sensory feedback is available. To this end, we have developed a method to model such actions using a hidden Markov model (HMM) representation. We proposed an approach to address two critical problems in action modeling: classifying human action-intent, and learning human skill, for which we elaborated on the method, procedure, and implementation issues in this paper. This work provides a framework for modeling and learning human actions from observations. The approach can be applied to intelligent recognition of manual actions and high-level programming of control input within a supervisory control paradigm, as well as automatic transfer of human skills to robotic systems.

The self-mobile space manipulator (SM/sup 2/) is a seven-DOF inspection robot, capable of walking along the pre-integrated I-beam truss structure of Space Station Freedom (SSF). The robot is capable of projecting cameras anywhere interior or exterior of the SSF, and will be an ideal tool for inspecting connectors, structures, and other facilities on the SSF. Using a full-scale model of a small segment of the SSF, experiments have been performed under two gravity compensation systems.

In space application, robot system are subject to unknown or unmodeled dynamics, for example, in the tasks of transporting an unknown payload or catching an unmodeled moving object. We discuss the parameterization problem in dynamic structure and adaptive control of a space robot system with an attitude-controlled base to which the robot is attached. We first derive the system kinematic and dynamic equations based on Lagrangian dynamics and the linear momentum conservation law. Based on the dynamic model developed, we discuss the problem of linear parameterization in term of dynamic parameters, and find that in joint space, the dynamics can be linearized by a set of combined dynamic parameters; however, in inertial space linear parameterization is impossible in general. Then we propose an adaptive control scheme in joint space, and present a simulation study to demonstrate its effectiveness and computational procedure. Because most takes are specified in inertial space instead of joint space, we discuss the issues associated to adaptive control in inertial space and identify two potential problem: unavailability of joint trajectory because the mapping from inertial space trajectory is dynamic-dependent and subject to uncertainty; and nonlinear parameterization in inertial space. We approach the problem by making use of the proposed joint space adaptive controller and updating the joint trajectory by the estimated dynamic parameters and given trajectory in inertial space. >

Robotic manipulators will play a significant role in the maintenance and repair of space stations and satellites, and other future space missions. Robot path planning and control for the above applications should be optimum, since any inefficiency in the planning may considerably risk the success of the space mission. This paper presents a global optimum path planning scheme for redundant space robotic manipulators to be used in such missions. In this formulation, a variational approach is used to minimize the objective functional. It is assumed that the gravity is zero in space, and the robotic manipulator is mounted on a completely free-flying base (spacecraft) and the attitude control (reaction wheels or thrust jets) is off. Linear and angular momentum conditions for this system lead to a set of mixed holonomic and nonholonomic constraints. These equations are adjoined to the objective functional using a Lagrange multiplier technique. The formulation leads to a system of differential and algebraic equations (DAEs). A numerical scheme for forward integration of this system is presented. A planar redundant space manipulator consisting of three arms and a base is considered to demonstrate the feasibility of the formulation. The approach to optimum path planning of redundant space robots is significant since most robots that have been developed for space applications so far are redundant. The kinematic redundancy of space robots offers efficient control and provides the necessary dexterity for extra-vehicular activity that exceeds human capacity. >

To perform simulations of partial or microgravity environments on Earth requires some method of compensation for the Earth's gravitational field. This paper discusses an active compensation system that modulates the tension in a counterweight support cable in order to minimize state deviation between the compensated body and the ideal weightless body. The system effectively compensates for inertial effects of the counterweight mass, viscous damping of all pulleys, and static friction in all parts of the gravity compensation (GC) system using a hybrid PI (proportional plus integral)/fuzzy control algorithm. The dynamic compensation of inertia and viscous damping is performed by PI control, while static friction compensation is performed by the fuzzy system. The system provides a very precise gravity compensation force, and is capable of nonconstant gravity force compensation in the case that the payload mass is not constant. >

In this paper, we discuss the problem of the switching surface for a multivariable variable structure system. We first propose a new switching surface by which robustness of the controlled system to disturbances is provided. Based on the switching surface, we develop control input to drive the system to track the desired trajectory. This control input does not produce the chattering that normally occurs when a conventional control input is used. We apply the proposed method to a multivariable linear system and discuss the simulation results to show the computational procedure and its effectiveness. >

The self-mobile space manipulator, (SM)/sup 2/, is a simple, 5-DOF, 1/3 scale, laboratory version of a robot designed to walk on the trusswork and other exterior surfaces of Space Station Freedom. It will be capable of routine tasks such as inspection, parts transportation and simple maintenance procedures. The authors have designed and built the robot and gravity compensation system to permit simulated zero-gravity experiments. The authors have developed the control system for the (SM)/sup 2/ including control hardware architecture and operating system, control station with various interfaces, hierarchical control structure, multi-phase control strategy for step motion, and various low-level controllers. The system provides operator friendly, real-time monitoring, robust control for 3-D locomotion movements of the flexible robot. A hierarchical structure allows the control to be executed in various levels autonomously or by teleoperation, and a multiphase control strategy facilitates the control in different tasks. Based on the dynamic model developed, a linear-structured joint-level controller and model-based control scheme with acceleration feedback is being implemented to provide a stable and fast motion. The configuration-independent control scheme allows the control parameters to adapt to changes in system dynamics due to the robot configuration variation. With a variety of low-level controllers which the authors developed, the system has demonstrated robustness to the uncertainties in modeling and in payload. >