IROS 2017 论文列表

One of the biggest challenges of humanoid robots is to keep the balance at any moment, as they can be subject to different types of external perturbations. Therefore, push recovery is a relevant issue in humanoid robotics, which still represents an open challenge. In literature, the most used method is based on the capture point, where a reduced model of the robot is used to compute recovery motions. In this paper we want to explore the problem with a different approach, by using whole-body models combined with optimal control. The robot is subject to an external perturbation, which is taken into account in the system dynamics as continuous external force. Optimal control is used to generate whole-body recovery motions by using the whole-body dynamic model of the robot. Instead of the capture point criterion, we optimize for a stable motion that allows to perform recovery within one step, at the end of which the robot is in a state of zero velocity. We apply the method to the model of the humanoid robot HeiCub, where serial springs can be mounted in the knee and ankle pitch actuators to change them into Series Elastic Actuators (SEA). The recovery motion is computed for both the model with and without SEA.

Optical coherence tomography (OCT) is an interferometric imaging modality with spatial resolution in the micrometer range. The OCT signal can be used to detect small structures to measure deformation or to characterize tissue. Moreover, OCT can be realized through a single optical fiber, i.e., it can be easily integrated with instruments. However, to use OCT for intra-operative guidance its spatial alignment needs to be established. Hence, we consider eye-in-hand calibration between the 1D OCT imaging and a 6D robotic position system. We present a method to perform pivot calibration for OCT and based on this introduce pivot+d, a new 1D to 6D eye-in-hand calibration. We provide detailed results on the convergence and accuracy of our method and use translational and rotational ground truth to show that our methods allow for submillimeter positioning accuracy of an OCT beam with a robot. For pivot calibration we observe a mean translational error of 0.5161 ± 0.4549 mm while pivot+d shows 0.3772 ± 0.2383 mm. Additionally, pivot+d improves rotation detection by about 8° when compared to pivot calibration.

Recently, image representations derived from Convolutional Neural Networks (CNNs) have been demonstrated to achieve impressive performance on a wide variety of tasks, including place recognition. In this paper, we take a step deeper into the internal structure of CNNs and propose novel CNN-based image features for place recognition by identifying salient regions and creating their regional representations directly from the convolutional layer activations. A range of experiments is conducted on challenging datasets with varied conditions and viewpoints. These reveal superior precision-recall characteristics and robustness against both viewpoint and appearance variations for the proposed approach over the state of the art. By analyzing the feature encoding process of our approach, we provide insights into what makes an image presentation robust against external variations.

Collaborative robots working on a common task are necessary for many applications. One of the challenges for achieving collaboration in a team of robots is mutual tracking and identification. We present a novel pipeline for online vision-based detection, tracking and identification of robots with a known and identical appearance. Our method runs in realtime on the limited hardware of the observer robot. Unlike previous works addressing robot tracking and identification, we use a data-driven approach based on recurrent neural networks to learn relations between sequential inputs and outputs. We formulate the data association problem as multiple classification problems. A deep LSTM network was trained on a simulated dataset and fine-tuned on small set of real data. Experiments on two challenging datasets, one synthetic and one real, which include long-term occlusions, show promising results.

Laser Guided Vehicles (LGV) are largely used in industrial contexts for the autonomous dispatching of huge loads. LGVs of automated warehouses are driven by supervisory systems which assign their tasks depending on the process demands. The vehicle workspace is partially structured and it is normally shared with several independent agents like other LGVs, or human operated forklifts, or humans which directly interact with the vehicle itself. For efficiency and for safety reasons, LGVs move along pre-defined paths so that, if an unexpected obstacle is detected, they must be promptly stopped so as to avoid impacts. As shown in the paper, the safety problem is normally handled by limiting the plant productivity. This is clearly an improper approach, which can be overcome by considering a novel velocity planning strategy. The one proposed in this paper guarantees high safety standards, thus preserving the physical integrity of human coworkers, by simultaneously improving the plant productivity.

In this paper, we propose an online learning algorithm based on a Rao-Blackwellized particle filter for spatial concept acquisition and mapping. We have proposed a nonparametric Bayesian spatial concept acquisition model (SpCoA). We propose a novel method (SpCoSLAM) integrating SpCoA and FastSLAM in the theoretical framework of the Bayesian generative model. The proposed method can simultaneously learn place categories and lexicons while incrementally generating an environmental map. Furthermore, the proposed method has scene image features and a language model added to SpCoA. In the experiments, we tested online learning of spatial concepts and environmental maps in a novel environment of which the robot did not have a map. Then, we evaluated the results of online learning of spatial concepts and lexical acquisition. The experimental results demonstrated that the robot was able to more accurately learn the relationships between words and the place in the environmental map incrementally by using the proposed method.

Robotic failure is all too common in unstructured robot tasks. Despite well-designed controllers, robots often fail due to unexpected events. Robots under a sense-plan-act paradigm do not have an additional loop to check their actions. In this work, we present a principled methodology to bootstrap online robot introspection for contact tasks. In effect, we seek to enable the robot to recognize and expect its behavior, else detect anomalies. We postulated that noisy wrench data inherently contains patterns that can be effectively represented by a vocabulary. The vocabulary is obtained by segmenting and encoding data. And when wrench information represents a sequence of sub-tasks, the vocabulary represents a set of words or sentence and provides a unique identifier. The grammar, which can also include unexpected events, was classified both offline and online for simulated and real robot experiments. Multi-class Support Vector Machines (SVMs) were used offline, while online probabilistic SVMs were used to give temporal confidence to the introspection result. Our work's contribution is the presentation of a generalizable online semantic scheme that enables a robot to understand its high-level state whether nominal or anomalous. It is shown to work in offline and online scenarios for a particularly challenging contact task: snap assemblies. We perform the snap assembly in one-arm simulated and real one-arm experiments and a simulated two-arm experiment. The data set itself is also fully available online and provides a valuable resource by itself for this type of contact task. Our verification mechanism can be used by high-level planners or reasoning systems to enable intelligent failure recovery or determine the next most optimal manipulation skill to be used. Supplemental information, code, data, and other supporting documentation can be found at [1].

The identification of inertial parameters is crucial to achieve high-performance model-based control of legged robots. The inertial parameters of the legs are typically not altered during expeditions and therefore are best identified offline. On the other hand, the trunk parameters depend on the modules mounted on the robot, like a motor to provide the hydraulic power, or different sets of cameras for perception. This motivates the use of recursive approaches to identify online mass and the position of the Center of Mass (CoM) of the robot trunk, when a payload change occurs. We propose two such approaches and analyze their robustness in simulation. Furthermore, experimental trials on our 80-kg quadruped robot HyQ show the applicability of our strategies during locomotion to cope with large payload changes that would otherwise severely compromise the balance of the robot.

Inverse dynamics models are applied to a plethora of robot control tasks such as computed-torque control, which are essential for trajectory execution. The analytical derivation of such dynamics models for robotic manipulators can be challenging and depends on their physical characteristics. This paper proposes a machine learning approach for modeling inverse dynamics and provides information about its implementation on a physical robotic system. The proposed algorithm can perform online multi-target learning, thus allowing efficient implementations on real robots. Our approach has been tested both offline, on datasets captured from three different robotic systems and online, on a physical system. The proposed algorithm exhibits state-of-the-art performance in terms of generalization ability and convergence. Furthermore, it has been implemented within ROS for controlling a Baxter robot. Evaluation results show that its performance is comparable to the built-in inverse dynamics model of the robot.

This paper presents an online technique which employs incremental support vector regression to learn the damping term of an underwater vehicle motion model, subject to dynamical changes in the vehicle's body. To learn the damping term, we use data collected from the robot's on-board navigation sensors and actuator encoders. We introduce a new sample-efficient methodology which accounts for adding new training samples, removing old samples, and outlier rejection. The proposed method is tested in a real-world experimental scenario to account for the model's dynamical changes due to a change in the vehicle's geometrical shape.

Human detection and tracking are essential aspects to be considered in service robotics, as the robot often shares its workspace and interacts closely with humans. This paper presents an online learning framework for human classification in 3D LiDAR scans, taking advantage of robust multi-target tracking to avoid the need for data annotation by a human expert. The system learns iteratively by retraining a classifier online with the samples collected by the robot over time. A novel aspect of our approach is that errors in training data can be corrected using the information provided by the 3D LiDAR-based tracking. In order to do this, an efficient 3D cluster detector of potential human targets has been implemented. We evaluate the framework using a new 3D LiDAR dataset of people moving in a large indoor public space, which is made available to the research community. The experiments analyse the real-time performance of the cluster detector and show that our online learned human classifier matches and in some cases outperforms its offline version.

Information gathering algorithms aim to intelligently select the robot actions required to efficiently obtain an accurate reconstruction of a physical process, such as an occupancy map, or a magnetic field. Many recent works have proposed algorithms for information gathering. However, these algorithms employ discretization of the state space, which makes them computationally intractable for robotic systems with complex dynamics. Moreover, most algorithms are not suited for online information gathering tasks. This paper presents a novel approach that tackles the two aforementioned issues. Specifically, our approach includes two intertwined steps: a Gaussian processes (GPs)-based prediction that allows a robot to identify highly unexplored locations, and an RRT*-based informative path planning that guides the robot towards those locations. The combination of the two steps allows an online realization of the algorithm, while eliminates the need of discretization. We demonstrate the effectiveness of the proposed algorithm in simulations, as well as with an experiment in which a ground-based robot explores the magnetic field intensity within an indoor environment populated with obstacles.

This paper describes an algorithm for visually detecting crop stalks in-situ. Field-based crop stalk detection is a challenging computer vision problem due to occlusion by leaves, similarity of color and texture between stalks and surrounding foliage, and high stalk density within rows. Detecting stalks for grasping adds further challenges such as computation time and computing hardware. The proposed algorithm utilizes multiple stereo images taken from an unmanned ground vehicle and exploits geometric features specific to stalks such as vertical continuity, height, and the direction of surface normals. The described hardware and software pipeline is capable of detecting stalks for a manipulator grasp in approximately 12 seconds, and preliminary results show that hardware improvements can reduce detection time to 4 seconds. The algorithm is tested on 378 point clouds generated from 916 images of Sorghum bicolor, a grain crop. Performance is evaluated according to two methods, demonstrating precision greater than 93%.

Modern consumer RGB-D cameras are affordable and provide dense depth estimates at high frame rates. Hence, they are popular for building dense environment representations. Yet, the sensors often do not provide accurate depth estimates since the factory calibration exhibits a static deformation. We present a novel approach to online depth calibration that uses a visual SLAM system as reference for the measured depth. A sparse map is generated and the visual information is used to correct the static deformation of the measured depth while missing data is extrapolated using a small number of thin plate splines (TPS). The corrected depth can then be used to improve the accuracy of the sparse RGB-D map and the 3D environment reconstruction. As more data becomes available, the depth calibration is updated on the fly. Our method does not rely on a planar geometry like walls or a one-to-one-pixel correspondence between color and depth camera. Our approach is evaluated in real-world scenarios and against ground truth data. Comparison against two popular self-calibration methods is performed. Furthermore, we show clear visual improvement on aggregated point clouds with our method.

Traditional path-planning involves (1) choosing start and goal points, (2) calculating a path, and (3) following that path. There are, however, many real world scenarios where an agent might need to change its goal and replan, which frequently includes expensively calculating a new path from scratch. We propose an adaptation to RRT∗ that locally rewires the RRT∗ tree as the robot moves and path costs change. Local rewiring takes advantage of information already existing in the tree and makes small adaptations that accommodate changes. Rewiring adds computational overhead during robot travel, but allows replanning in real time with approximately constant overhead. Empirical studies demonstrate that computational costs are lower than alternative replanners in 2D worlds with moderate obstacle density, and the resulting paths approach optimality as more time is allowed to replan.

In this paper, we propose a GPU parallelized SLAM system capable of using photometric and inertial data together with depth data from an active RGB-D sensor to build accurate dense 3D maps of indoor environments. We describe several extensions to existing dense SLAM techniques that allow us to operate in real-time onboard memory constrained robotic platforms. Our primary contribution is a memory management algorithm that scales to large scenes without being limited by GPU memory resources. Moreover, by integrating a visual-inertial odometry system, we robustly track the camera pose even on an agile platform such as a quadrotor UAV. Our robust camera tracking framework can deal with fast camera motions and varying environments by relying on depth, color and inertial motion cues. Global consistency is achieved via regular checking for loop closures in conjunction with a pose graph, as a basis for corrective deformation of the 3D map. Our efficient SLAM system is capable of producing highly dense meshes up to 5mm resolution at rates close to 60Hz fully onboard a UAV. Experimental validations both in simulation and on a real-world platform, show that our approach is fast, more robust and more memory efficient than state-of-the-art techniques, while obtaining better or comparable accuracy.

In this paper, we propose a part-based particle filter for on-road vehicle tracking. The proposed model takes part-based strategies into account in a particle filter. By introducing a hidden state vehicle center position, vehicle parts particles can be updated efficiently as a whole sharing same motion. With a pre-trained appearance and geometric model, tracker can distinguish parts with rich information from invalid parts to make a more precise prediction. Meanwhile some priori knowledge about the moving pattern of vehicles in well-structured on-road environment is learned, and can be used in the inference of measurement model and motion model to improve tracking performance and efficiency. Experiments were conducted with real data collected in Beijing to examine the performance in different situations on both the advantages and challenges. The Beijing highway dataset for on-road vehicle tracking will be opened to the society. We compare our method with the state-of-the-art approaches. Result demonstrate that the proposed algorithm are able to handle occlusion and aspect ratio change in on-road vehicle tracking problem.

Capturing realistic human behaviors is essential to learn human models that can later be transferred to robots. Recent improvements in virtual reality (VR) head-mounted displays provide a viable way to collect natural examples of human behavior without the difficulties often associated with capturing performances in a physical environment. We present a realistic, cluttered, VR environment for experimentation with household tasks paired with a semantic extraction and reasoning system able to utilize data collected in real-time and apply ontology-based reasoning to learn and classify activities performed in VR. The system performs continuous segmentation of the motions of users' hands and simultaneously classifies known actions while learning new ones on demand. The system then constructs a graph of all related activities in the environment through its observations, extracting the task space utilized by observed users during their performance. The action recognition and learning system was able to maintain a high degree of accuracy of around 92% while dealing with a more complex and realistic environment compared to earlier work in both physical and virtual spaces.

This paper presents an adaptive side-by-side human-robot companion approach for navigation in urban dynamic environments, based on the anticipative kinodynamic planning. The adaptive means that the robot is capable of adjusting its motion to the behavior of the person being accompanied. Our main objective is to optimize in real time the path performed by the pair human-robot, by modifying dynamically the angle and distance between both throughout different locations of the path. We have defined a new cost function for finding the best planned path that takes into account the cost of the geometrical configuration between the human and the robot. Moreover, we have modified the Extended Social Force Model (SFM) to include the required forces to maintain the angle and distance between the robot and human while the human-robot pair is moving towards the shared goal. The method has been validated throughout a large set of simulations and real-live experiments.