IROS 2016 Papers

Module localization is an important aspect of the operation of self-reconfigurable robots. The knowledge of spatial positions of modules, or at least of the overall shape which the modules form, is the usual prerequisite for reconfiguration planning. We present a general, decentralized algorithm for determining the positions of modules placed on a cubic grid from local sensor information. The connection topology of the robot is arbitrary. We assume that a module can sense the presence of its immediate neighbors on the grid and determine their positions in its own local coordinate system, but cannot sense the orientations of the coordinate systems of its neighbors. Since orientation cannot be directly communicated between modules, the modules can only exchange information about the relative positions of their neighbors. The algorithm aggregates this information over the entire network of modules and narrows down the set of valid positions for each module as far as possible. If there exists a unique locally-consistent assignment of coordinates to all modules then it is found.

This paper explores the problem of building a structure of a desired shape, using re-usable interlocking blocks. Blocks are cubes; we make use of nine different types of cubes, each with different arrangements of male and female connectors on the six sides of the cube. The desired shape is specified by a set of voxels. We propose an algorithm that lays out cubes in a particular pattern to give the desired shape and gives a fairly simple assembly order.

Eye tracking technology has shown promising results for allowing hands-free control of robotically-mounted cameras and tools. However existing systems present only limited capabilities in allowing the full range of camera motions in a safe, intuitive manner. This paper introduces a framework for the recognition of surgeon intention, allowing activation and control of the camera through natural gaze behaviour. The system is resistant to noise such as blinking, while allowing the surgeon to look away safely at any time. Furthermore, this paper presents a novel approach to control the translation of the camera along its optical axis using a combination of eye tracking and stereo reconstruction. Combining eye tracking and stereo reconstruction allows the system to determine which point in 3D space the user is fixating, enabling a translation of the camera to achieve the optimal viewing distance. In addition, the eye tracking information is used to perform automatic laser targeting for laser ablation. The desired target point of the laser, mounted on a separate robotic arm, is determined with the eye tracking thus removing the need to manually adjust the laser's target point before starting each new ablation. The calibration methodology used to obtain millimetre precision for the laser targeting without the aid of visual servoing is described. Finally, a user study validating the system is presented, showing clear improvement with median task times under half of those of a manually controlled robotic system.

The complexity of robotics comes from the tight interactions between hardware, complex softwares, and environments. While real world experience is the only way to assess the efficiency and robustness of a robotics system, simulations help to pave the way to actual experiments. But an overall robotics system requires simulations at a level of realism which no holistic simulator can provide, given the wide spectrum of disciplines and physical processes involved. This paper presents a way to integrate various simulators, in a distributed, scalable and repeatable way, to benefit from their different advantages and get the best fitted and accurate simulation for a given robotics system. It depicts how the MORSE open-source robotics simulator is adapted to comply with the High Level Architecture standard, thus allowing the reuse of numerous dedicated realistic simulators. Two examples of the integration of simulators are provided.

The information used to determine the internal model of a robot system emerges from individual interactions with the environment. Knowledge about a specific internal model can be acquired by means of model learning techniques based on supervised machine learning tools, such as neural networks. One of the main challenges is to specify the starting point for the learning process. We propose a methodology for initial weight estimation based on the robot morphology. We use the generation of saccades in a robotic head as a case study to evaluate the performance of this approach. Our results suggest that this methodology, based on learning the relationship between the morphological parameters of the robot and those in its internal model, improves the quality of weight initialization, resulting in a considerable speed-up in the process of learning the internal model.

We consider an autonomous exploration problem in which a mobile robot is guided by an information-based controller through an a priori unknown environment, choosing to collect its next measurement at the location estimated to be most informative within its current field of view. We propose a novel approach to predict mutual information (MI) using Bayesian optimization. Over several iterations, candidate sensing actions are suggested by Bayesian optimization and added to a committee that repeatedly trains a Gaussian process (GP). The GP estimates MI throughout the robot's action space, serving as the basis for an acquisition function used to select the next candidate. The best sensing action in the committee is executed by the robot. This approach is compared over several environments with two batch methods, one which chooses the most informative action from a set of pseudo-random samples whose MI is explicitly evaluated, and one that applies GP regression to this sample set. Our computational results demonstrate that the proposed method provides not only computational efficiency and rapid map entropy reduction, but also robustness in comparison with competing approaches.

Much of estimation of human internal state (goal, intentions, activities, preferences, etc.) is passive: an algorithm observes human actions and updates its estimate of human state. In this work, we embrace the fact that robot actions affect what humans do, and leverage it to improve state estimation. We enable robots to do active information gathering, by planning actions that probe the user in order to clarify their internal state. For instance, an autonomous car will plan to nudge into a human driver's lane to test their driving style. Results in simulation and in a user study suggest that active information gathering significantly outperforms passive state estimation.

Here we experimentally quantify the effects of wing morphological and inertial parameters on flapping flight performance. Through running at-scale, passive pitching experiments with different wing designs, we compare the relative importance of wing inertia, wing shape, and wing-actuation pairing. We find wing inertia strongly influences the coupling between stroke and pitch dynamics, which directly impacts lift production and efficiency. Flapping resonance frequency is reduced as wing aspect ratio or area moment increases. Further, wing leading edge design strongly influences chordwise center of pressure, which further impacts pitching dynamics. Based on our experimental results we propose a new wing design and measure 37% increase in mean lift relative to a previous work [1].

This paper presents a method which allows robots to infer a human's hierarchical intent from partially observed RGBD videos by imagining how the human will behave in the future. This capability is critical for creating robots which can interact socially or collaboratively with humans. We represent intent as a novel hierarchical, compositional, and probabilistic And-Or graph structure which describes a relationship between actions and plans. We infer human intent by reverse-engineering a human's decision-making and action planning processes under a Bayesian probabilistic programming framework. We present experiments from a 3D environment which demonstrate that the inferred human intent (1) matches well with human judgment, and (2) provides useful contextual cues for object tracking and action recognition.

We address the problem of extracting human body posture labels, upper body orientation and the spatial location of individual body parts from near-infrared (NIR) images depicting patterns of retro-reflective markers. The analyzed patterns originate from the observation of humans equipped with protective high-visibility garments that represent common safety equipment in the industrial sector. Exploiting the shape of the observed reflectors we adopt shape matching based on the chamfer distance and infer one of seven discrete body posture labels as well as the approximate upper body orientation with respect to the camera. We then proceed to analyze the NIR images on a pixel scale and estimate a figure-ground segmentation together with human body part labels using classification of densely extracted local image patches. Our results indicate a body posture classification accuracy of 80% and figure-ground segmentations with 87% accuracy.

In this paper we give an evaluation of different direct methods for computing frame-to-frame motion estimates of a moving sensor rig composed of an RGB-D camera and an inertial measurement unit. In particular, we compare how semi-dense and fully dense tracking methods, with and without the aid of an inertial measurement unit (IMU), perform with respect to changes in image resolution, shutter speed, frame-rates, as well as image and depth noise. To perform an accurate and unbiased evaluation we employ a series of synthetically generated datasets using a simulated sensor rig composed of an RGB-D camera and an IMU. Our findings show that in the absence of motion blur or for cameras with high enough frame-rates relative to the camera motion, the methods are comparable when taking in consideration both accuracy and computation time.

We present an architecture for online, incremental scene modeling which combines a SLAM-based scene understanding framework with semantic segmentation and object pose estimation. The core of this approach comprises a probabilistic inference scheme that predicts semantic labels for object hypotheses at each new frame. From these hypotheses, recognized scene structures are incrementally constructed and tracked. Semantic labels are inferred using a multi-domain convolutional architecture which operates on the image time series and which enables efficient propagation of features as well as robust model registration. To evaluate this architecture, we introduce a large-scale RGB-D dataset JHUSEQ-25 as a new benchmark for the sequence-based scene understanding in complex and densely cluttered scenes. This dataset contains 25 RGB-D video sequences with 100,000 labeled frames in total. We validate our method on this dataset and demonstrate improved performance of semantic segmentation and 6-DoF object pose estimation compared with methods based on the single view.

Real-time outdoor navigation in highly dynamic environments is an crucial problem. The recent literature on real-time static SLAM don't scale up to dynamic outdoor environments. Most of these methods assume moving objects as outliers or discard the information provided by them. We propose an algorithm to jointly infer the camera trajectory and the moving object trajectory simultaneously. In this paper, we perform a sparse scene flow based motion segmentation using a stereo camera. The segmented objects motion models are used for accurate localization of the camera trajectory as well as the moving objects. We exploit the relationship between moving objects for improving the accuracy of the poses. We formulate the poses as a factor graph incorporating all the constraints. We achieve exact incremental solution by solving a full nonlinear optimization problem in real time. The evaluation is performed on the challenging KITTI dataset with multiple moving cars. Our method outperforms the previous baselines in outdoor navigation.

We provide a novel incremental data association method to complement our previous work on acoustic structure from motion (ASFM), which recovers 3D scene structure from multiple 2D sonar images, while at the same time localizing the sonar. Given point features extracted from multiple overlapping sonar images, our algorithm automatically finds the correspondences between the features. Our data association method uses information about the geometric correlations of the entire set of landmarks to reject spurious measurements or false positives that might otherwise have been accepted. For each new sonar measurement, the algorithm uses a gating procedure to narrow the landmark match search space. Using the pruned surviving candidate correspondences, we identify the correct hypothesis based on a posterior compatibility cost, penalizing for null matches to avoid all measurements being declared new landmarks. Unlike other methods, ASFM does not require any planar scene assumptions and uses constraints from more than two images to increase accuracy in both mapping and localization. We evaluate our algorithm in simulation and demonstrate successful data association results on real sonar images.

Most modern approaches for tactile object recognition with robotic hands do not use proprioceptive data. In those that do, a limited number of objects with similar shapes is recognized. Furthermore, Self-Organizing Maps (SOM) based on raw values of joint angles/torques are frequently implemented which requires large sets of training data. In this paper, we present an approach based only on joint angles of a robotic hands to identify the shape of an object regardless its size and position within the hand. A representation of the joint angles is created to endow the robotic hand with proprioception. Support Vector Machine (SVM) is implemented for shape identification using patterns or signatures generated on this representation when the objects are grasped. To illustrate the scope of this method, tests are performed on five shapes present in common objects. Both SVM and SOM trained with signatures were at least 10% more accuracy than the ones trained with raw values of joint angles. Training sets are reduced at least 85% with respect to other works. An accuracy of 94% was obtained on large ranges of dimensions of the shapes and positions.

This paper presents the robotic hammering task performed by single arm and dual arm. In order to analyze optimal hammering task, the analytical model of external and internal impulses are taken as the measure. Desired hammering task implies that the hammer simultaneously exerts maximum external impulse on nail and minimum internal impulses on robot or human joints. The advantages of dual arm manipulator with muscles over dual arm without muscles and single arm model with muscle and without muscle are presented. Moreover, the effective mass in nailing operation is experimentally obtained for two different materials and it is integrated into the external impulse model of the hammering task. Furthermore, the hammering workspace is optimized by considering the minimum internal and maximum external impulses.

Estimating Cartesian contact forces and torques enables external force supervision for robotic manipulators and even force-controlled applications while avoiding the need for additional external sensing. Redundant manipulators facilitate the problem of Cartesian contact force and torque estimation (CCFE) at the TCP, since an increased amount of joint level information is available for estimating the six components of external wrench. In this paper, we point out that the errors in CCFE estimates arising from inevitable uncertainties in available joint level information (either measurements or estimates) depend on the manipulator configuration. Based on this, an algorithm is proposed that resolves redundancy in an optimal way with respect of the achievable CCFE quality. For a given TCP pose, the joint configuration resulting in minimized CCFE errors is calculated. The proposed method is verified by means of experimental data gathered from a 7DOF manipulator.

Conventionally, a cable driven parallel mechanism (CDPM) pose is obtained through the forward kinematics from measurements of the cable lengths. However, this estimation method can be limiting for some applications requiring more precision. This paper proposes to use cable angle position sensors in addition to cable length measurements in order to improve the accuracy of such mechanisms. The robot pose is first obtained individually by the cable length measurements and the cable angle position measurements. A data fusion scheme combining these two types of measurements is then proposed in order to improve the CPDM accuracy. Finally, simulations and experiments are presented in order to assess the benefits of using cable angle position sensors on the CDPM.