IROS 2015 Papers

In this paper, we present a monocular visual-inertial odometry algorithm which, by directly using pixel intensity errors of image patches, achieves accurate tracking performance while exhibiting a very high level of robustness. After detection, the tracking of the multilevel patch features is closely coupled to the underlying extended Kalman filter (EKF) by directly using the intensity errors as innovation term during the update step. We follow a purely robocentric approach where the location of 3D landmarks are always estimated with respect to the current camera pose. Furthermore, we decompose landmark positions into a bearing vector and a distance parametrization whereby we employ a minimal representation of differences on a corresponding σ-Algebra in order to achieve better consistency and to improve the computational performance. Due to the robocentric, inverse-distance landmark parametrization, the framework does not require any initialization procedure, leading to a truly power-up-and-go state estimation system. The presented approach is successfully evaluated in a set of highly dynamic hand-held experiments as well as directly employed in the control loop of a multirotor unmanned aerial vehicle (UAV).

In this paper, we present an appearance-based visual SLAM approach that focuses on detecting loop closures across seasons. Given two image sequences, our method first extracts one descriptor per image for both sequences using a deep convolutional neural network. Then, we compute a similarity matrix by comparing each image of a query sequence with a database. Finally, based on the similarity matrix, we formulate a flow network problem and compute matching hypotheses between sequences. In this way, our approach can handle partially matching routes, loops in the trajectory and different speeds of the robot. With a matching hypothesis as loop closure information and the odometry information of the robot, we formulate a graph based SLAM problem and compute a joint maximum likelihood trajectory.

This paper presents a novel control method to stabilize the whole-body motion of humanoid robots when climbing vertical ladders and transitioning between ladders and catwalks. In such environments, the body of the robot tends to incline and rotate because of the slippery surfaces. The inclination and rotation may cause the robot to fail to grasp and thus collide with the rungs. The proposed method modifies the subsequent contact position in real time based on the error of the current robot posture estimated with inertial measurement units (IMUs) and actual joint angles. This paper also presents a method of generating motion by minimizing the contact wrench. This method satisfies hardware limitations, such as collision avoidance, joint torque limits, and joint limits. Applying these methods to a humanoid robot, we realize the robust climbing and descending of multiple rungs of a vertical ladder and bidirectional transitioning from ladders to catwalks.

We present an algorithm for generating open-loop trajectories that solve the problem of rearrangement planning under uncertainty. We frame this as a selection problem where the goal is to choose the most robust trajectory from a finite set of candidates. We generate each candidate using a kinodynamic state space planner and evaluate it using noisy rollouts. Our key insight is we can formalize the selection problem as the “best arm” variant of the multi-armed bandit problem. We use the successive rejects algorithm to efficiently allocate rollouts between candidate trajectories given a rollout budget. We show that the successive rejects algorithm identifies the best candidate using fewer rollouts than a baseline algorithm in simulation. We also show that selecting a good candidate increases the likelihood of successful execution on a real robot.

Materials and textures identification is a desired ability for robots. Developing such systems require tactile sensors that have enough sensitivity and spatial resolution, and the computational intelligence to meaningfully interpret sensor data. This paper introduces a texture classification algorithm utilizing support vector machine (SVM) classifier. Data taken from a novel three axis tactile sensor that utilize magnetic flux measurements for transduction was used to obtain the three dimensional tactile data. Frobenius norm calculated from the covariance matrix of the above data and the mean values of the three dimensional sensor data were used as features. Palpation velocity and small vertical load variances had minimum influence on the proposed algorithm. We have compared this algorithm with two other classification methods. They are: classify using the feature spatial period that is calculated from principal frequencies of the textures/material, and classify using neural network classifier with special properties of each material's tactile signals as features. For eight classes of material, the proposed algorithm performed faster and more accurately than the comparators when the scanning velocity and the vertical load varied.

This paper proposes a new robust position control method for a novel series elastic actuator (SEA). It is a wellknown fact that SEAs provide many benefits in force control, e.g., lower reflected inertia and impedance, greater shock tolerance, safety, and so on. However, current SEA designs have a common performance limitation due to the compromise on the selection of spring stiffness. The performance of an SEA can be significantly improved by changing the stiffness of the spring; however, designing a variable-stiffness SEA is a quite challenging task. In this paper, a novel variable-stiffness SEA, which can relax the fundamental performance limitation of conventional SEAs, is proposed. It consists of torsional and linear springs, which have different compliances, in series. The soft and hard springs improve the performance when low and high force control applications are performed, respectively. Although SEAs have several advantages in force control, their position control problem is more complicated than the force control one. Moreover, using extra springs increases the number of vibration mode, which may significantly deteriorate the performance, in the position control problem of the SEA. In this paper, a new position control system, which improves the performance by increasing the robustness and suppressing the vibration, is proposed for a novel SEA. Experimental results are given to validate the proposal.

Robots based on tensegrity structures have the potential to be robust, efficient and adaptable. While traditionally being difficult to control, recent control strategies for ball-shaped tensegrity robots have successfully enabled punctuated rolling, hill-climbing and obstacle climbing. These gains have been made possible through the use of machine learning and physics simulations that allow controls to be “learned” instead of being engineered in a top-down fashion. While effective in simulation, these emergent methods unfortunately give little insight into how to generalize the learned control strategies and evaluate their robustness. These robustness issues are especially important when applied to physical robots as there exists errors with respect to the simulation, which may prevent the physical robot from actually rolling. This paper describes how the robustness can be addressed in three ways: 1) We present a dynamic relaxation technique that describes the shape of a tensegrity structure given the forces on its cables; 2) We then show how control of a tensegrity robot “ball” for locomotion can be decomposed into finding its shape and then determining the position of the center of mass relative to the supporting polygon for this new shape; 3) Using a multi-step Monte Carlo based learning algorithm, we determine the structural geometry that pushes the center of mass out of the supporting polygon to provide the most robust basic mobility step that can lead to rolling. Combined, these elements will give greater insight into the control process, provide an alternative to the existing physics simulations and offer a greater degree of robustness to bridge the gap between simulation and hardware.

Incorrect landmark and loop closure measurements can cause standard SLAM algorithms to fail catastrophically. Recently, several SLAM algorithms have been proposed that are robust to loop closure errors, but it is shown in this paper that they cannot provide robust solutions when landmark measurement errors occur. The root cause of this problem is that the robust SLAM algorithms only focus on generating solutions that are locally consistent (i.e. each measurement agrees with its corresponding estimates) rather than globally consistent (i.e. all of the measurements in the solution agree with each other). Moreover, these algorithms do not attempt to maximize the number of correct measurements included in the solution, meaning that often correct measurements are ignored and the solution quality suffers as a result. This paper proposes a new formulation of the robust SLAM problem that seeks a globally consistent map that also maximizes the number of measurements included in the solution. In addition, a novel incremental SLAM algorithm, called incremental SLAM with consistency-checking, is developed to solve the new robust SLAM problem. Finally, simulated and experimental results show that the new algorithm significantly outperforms state-of-the-art robust SLAM methods for datasets with incorrect landmark measurements and can match their performance for datasets with incorrect loop closures.

Moving objects within the hand is challenging, especially if the objects are of various shape and size. In this paper we use machine learning to learn in-hand manipulation of such various sized and shaped objects. The TWENDY-ONE hand is used, which has various properties that makes it well suited for in-hand manipulation: a high number of actuated joints, passive degrees of freedom and soft skin, six-axis force/torque (F/T) sensors in each fingertip, and distributed tactile sensors in the skin. A dataglove is used to gather training samples for teaching the required behavior. The object size information is extracted from the initial grasping posture. After training a neural network, the robot is able to manipulate objects of untrained sizes and shape. The results show the importance of size and tactile information. Compared to interpolation control, the adaptability for the initial posture gap could be greatly extended. Final results show that with deep learning the number of required training sets can be drastically reduced.

Recent developments in human-robot interaction brings about higher requirements for robot navigation. Existing Simultaneous Localization and Mapping (SLAM) algorithms face open challenges for navigation in complex dynamic environments due to presumptions of static environments or exceeding computational limitations. In this paper, we propose a robust graph SLAM formuation exploring Expectation Maximization algorithms to characterize landmark mobility while establishing the estimations of robot trajectory and the map. We evaluate the performance of existing robust SLAM algorithms as baselines, and validate the improvement of our new framework against datasets of dynamic environments with moving landmarks.

We consider how to directly extract a road map (also known as a topological representation) of an initially-unknown 2-dimensional environment via an on-line procedure which robustly computes a retraction of its boundaries. While such approaches are well known for their theoretical elegance, computing such representations in practice is complicated when the data is sparse and noisy. In this paper we present the online construction of a topological map and the implementation of a control law for guiding the robot to the nearest unexplored area. The proposed method operates by allowing the robot to localize itself on a partially constructed map, calculate a path to unexplored parts of the environment (frontiers), compute a robust terminating condition when the robot has fully explored the environment, and achieve loop closure detection. The proposed algorithm results in smooth safe paths for the robot's navigation needs. The presented approach is an any-time-algorithm which allows for the active creation of topological maps from laser-scan data, as it is being acquired. The resulting map is stable under variations to noise and the initial conditions. The key idea is the use of a flux-based skeletonization algorithm on the latest occupancy grid map. We also propose a navigation strategy based on a heuristic where the robot is directed towards nodes in the topological map that open to empty space. The method is evaluated on both synthetic data and in the context of active exploration using a Turtlebot 2. Our results demonstrate complete mapping of different environments with smooth topological abstraction without spurious edges.

We propose a novel flow control method to realize flow regulation and valveless flow switching in a microfluidic network. The proposed controller enables precise regulation of flow rates (or velocity fields) in micro-channels regardless of model uncertainties such as cross-sectional area and microchannel configuration. The controller can also minimize the pressure exerted on the microfluidic chip. The proposed system can dramatically improve performances of flow regulation and can overcome the limitations of conventional pump systems for microfluidic devices.

In this paper, a robust controller is designed to achieve accurate positioning of an unactuated surface vessel by using multiple unidirectional tugboats. After initially locating opposing tugboats to specific configurations, the control problem is transformed into a second order system with an uncertain non-symmetric input gain matrix. Upon applying a matrix decomposition, a robust controller is proposed. Detailed stability analysis ensured asymptotic tracking. Numerical simulation results demonstrate the efficiency of the proposed controller.

This paper introduces a novel concept of an air bearing test bench for CubeSat ground testing together with the corresponding dynamic parameter identification method. Contrary to existing air bearing test benches, the proposed concept allows three degree-of-freedom unlimited rotations and minimizes the influence of the test bench on the tested CubeSat. These advantages are made possible by the use of a robotic wrist which rotates air bearings in order to make them follow the CubeSat motion. Another keystone of the test bench is an accurate balancing of the tested CubeSat. Indeed, disturbing factors acting on the satellite shall be minimized, the most significant one being the gravity torque. An efficient balancing requires the CubeSat center of mass position to be accurately known. Usual techniques of dynamic parameter identification cannot be directly applied because of the frictionless suspension of the CubeSat in the test bench and, accordingly, due to the lack of external actuation. In this paper, a new identification method is proposed. This method does not require any external actuation and is based on the sampling of free oscillating motions of the CubeSat mounted on the test bench.

Paper folding is one of the most difficult tasks for multi-fingered robot hands because paper is deformable and its stiffness distribution is nonuniform. In this study, we aimed to achieve dexterous paper folding by extracting some dynamic motion primitives. Each primitive contains visual or force information, a physical model of a sheet of paper is used for analyzing its deformation, and a machine learning method is used for predicting its future state. We also propose a strategy for achieving valley folds in a sheet of paper twice in a row. One problem faced in folding paper is that, in the second fold, the crease line of the first fold disturbs the folding accuracy. We propose some new manipulation techniques to solve this problem. Finally, we show some demonstrations of paper folding achieved with a high success rate.

Realizing accurate robotic manipulation of flexible objects is very challenging. It asks for dexterous manipulation skill, and demands good adaptation to dynamic uncertainties due to deformation. Unlike traditional methods that basically depend on complex modeling, we follow the opposite approach by exploring possible simplifications based on high-speed motion strategy as well as high-speed visual sensing. As an application task concerned with robotic needle threading, this work demonstrates our philosophy by proposing a novel and straightforward high-speed strategy to speed up the manipulation while solves the issue of deformation uncertainty at the same time. The methodology is based on a very simple physical fact, namely that by rotating a thread with sufficient velocity, the thread can be approximated as a rigid object caused by the constantly applied centrifugal force during the rotation. Therefore, the needle threading problem is converted into a simpler peg-insertion problem, and the complexity of interaction between robot and deformable thread is significantly simplified. Based on the proposed strategy, simple technical methods then become possible for adoption to accomplish the challenging task. Experiments show that the proposed approach allows the needle threading task to be realized very rapidly.

This paper develops a novel robotic joint mechanism by means of twisting a small-diameter round-belt, which enables slow movements of joint rotation unlike direct-drive actuator mechanisms. The actuator mechanism proposed in this manuscript is composed of two small-diameter round-belts located near the joint with opposite configurations. The two round-belts are twisted by DC motors placed on a motor stage, which can be activated by a step motor. This novel mechanism realizes movement of the robotic joint around its axis due to contraction forces generated by twisting both round-belts. That is, these round-belts act as agonist and antagonist actuators for the robotic joint, which give a human-like compliance to the rotational motion of the joint. Experimental results using the proposed one-link robot show that a high position resolution of the joint control can be achieved by increases and decreases in the amount of twisting. In addition, we clearly indicate a linear relationship between twisting and the joint angle of the robot. Finally, this paper reveals that the antagonistic twist-actuator system has a secondary role as a gear reducer that is capable of largely decreasing the velocity of joint movement.

Several robotic systems have been proposed for removing blood from the brain in patients who have undergone a hemorrhagic stroke. In this paper we explore the use of imagebased feedback to address tissue deformation when aspirating a hemorrhage in a phantom model. This is the first time intraoperative image feedback has been used with a concentric tube robot in this application. We describe a layer by layer approach to motion planning. Computed tomography (CT) images are collected periodically during hemorrhage removal. After each CT scan, the robot's tip path is re-planned to account for the tissue deformation that has occurred since the previous scan. We compare open loop hemorrhage removal to our sequential imaging-replanning approach, illustrating that the latter has the potential to enhance the safety and efficacy of the procedure.

This paper presents a novel vision-based underwater robotic system for the identification and control of Crown- Of-Thorns starfish (COTS) in coral reef environments. COTS have been identified as one of the most significant threats to Australia's Great Barrier Reef. These starfish literally eat coral, impacting large areas of reef and the marine ecosystem that depends on it. Evidence has suggested that land-based nutrient runoff has accelerated recent outbreaks of COTS requiring extensive use of divers to manually inject biological agents into the starfish in an attempt to control population numbers. Facilitating this control program using robotics is the goal of our research. In this paper we introduce a vision-based COTS detection and tracking system based on a Random Forest Classifier (RFC) trained on images from underwater footage. To track COTS with a moving camera, we embed the RFC in a particle filter detector and tracker where the predicted class probability of the RFC is used as an observation probability to weight the particles, and we use a sparse optical flow estimation for the prediction step of the filter. The system is experimentally evaluated in a realistic laboratory setup using a robotic arm that moves a camera at different speeds and heights over a range of real-size images of COTS in a reef environment.