IROS 2018 论文列表

Kinematics-based collision detection is important for robot motion planning in unstructured terrain. Especially, planetary rovers require such capability as a single collision may lead to the termination of a mission. For onboard computation, typical numeric approaches are unsuitable as they are computationally expensive and unstable on rocky terrain; instead, a light-weight analytic solution (ACE: Approximate Clearance Evaluation) is planning to be used for the Mars 2020 rover mission. ACE computes the state bounds of articulated suspension systems from terrain height bounds, and assess the safety by checking the constraint violation of states with the worst-case values. ACE's conservative safety check approach can sometimes lead to over-pessimism: feasible states are often reported as infeasible, thus resulting in frequent false positive detection. In this paper, we introduce a computationally efficient probabilistic variant of ACE (called p-ACE) which estimates the probability distributions of states in real time. The advantage of having probability distributions over states, instead of deterministic bounds, is to provide more flexible and less pessimistic worst-case evaluation with probabilistic safety guarantees. Empirically derived distribution models are used to compute the total probability of constraint satisfaction, which is then used for path assessment. Through experiments with a high-fidelity simulator, we empirically show that p-ACE relaxes the deterministic state bounds without losing safety guarantees.

This paper presents a probabilistic approach for online dense reconstruction using a single monocular camera moving through the environment. Compared to spatial stereo, depth estimation from motion stereo is challenging due to insufficient parallaxes, visual scale changes, pose errors, etc. We utilize both the spatial and temporal correlations of consecutive depth estimates to increase the robustness and accuracy of monocular depth estimation. An online, recursive, probabilistic scheme to compute depth estimates, with corresponding covariances and inlier probability expectations, is proposed in this work. We integrate the obtained depth hypotheses into dense 3D models in an uncertainty-aware way. We show the effectiveness and efficiency of our proposed approach by comparing it with state-of-the-art methods in the TUM RGB-D SLAM & ICL-NUIM dataset. Online indoor and outdoor experiments are also presented for performance demonstration.

In this paper, we propose a probabilistic collision threat assessment algorithm for autonomous driving at road intersections that assesses a given traffic situation at an intersection reliably and robustly for an autonomous vehicle to cross the intersection safely, even in the face of violation vehicles (that is, vehicles in violation of traffic rules at the intersection). To this end, the proposed algorithm employs a detailed digital map to predict future paths of observed vehicles and then utilizes the predicted future paths to identify potential threats (vehicles) and potential collision areas, regardless of whether observed vehicles are obeying traffic rules at the intersection. Next, by means of Bayesian networks and time window filtering under an independent and distributed reasoning structure, it assesses the potential threats regarding the possibility of collision reliably and robustly, even under uncertain and incomplete noise data. Then, it has been tested and evaluated through in-vehicle testing on a closed urban test road under traffic conditions inclusive of non-violation and violation vehicles. In-vehicle testing results show that the performance of the proposed algorithm is sufficiently reliable to be used in decision-making for autonomous driving at intersections in terms of reliability and robustness, even in the face of violation vehicles.

Successful human-robot collaboration depends on a shared understanding of task state and current goals. In nonlinear or freeform tasks without an explicit task model, robot partners are unable to provide assistance without the ability to translate perception into meaningful task knowledge. In this paper, we explore the utility of multimodal recurrent neural networks (RNNs) with long short-term memory (LSTM) units for real-time subtask recognition in order to provide context-aware assistance during generic assembly tasks. We train RNNs to recognize specific subtasks in individual modalities, then combine the high-level representations of these networks through a nonlinear connection layer to create a multimodal subtask recognition system. We report results from implementing the system on a robot that uses the subtask recognition system to provide predictive assistance to a human partner during a laboratory experiment involving a human-robot team completing an assembly task. Generalizability of the system is evaluated through training and testing on separate tasks with some similar subtasks. Our results demonstrate the value of such a system in providing assistance to human partners during a freeform assembly scenario and increasing humans' perception of the robot's agency and usefulness.

We propose a collision avoidance method that incorporates the interactive behavior of agents and is proactive in dealing with the uncertainty of the future behavior of obstacles. The proposed method considers interactions that will be experienced by an autonomous surface vessel (ASV) in an environment governed by the international regulations for preventing collisions at sea (COLREGs). Our approach aims at encouraging dynamic obstacles to cooperate according to COLREGs. Therefore, we propose a strategy for assessing the cooperative behavior of obstacles, and the result of the assessment is used to adapt collision avoidance decisions within the Reciprocal Velocity Obstacles (RVO) framework. Moreover, we propose a predictive approach to solving known limitations of the RVO framework, and we present computationally feasible extensions that enable the use of complex dynamic models and objectives suitable for ASVs. We demonstrate the performance and potentials of our method through a simulation study, and the results show that the proposed method leads to proactive and more predictable ASV behavior compared with both Velocity Obstacles (VO) and RVO, especially when obstacles cooperate by following COLREGs.

Force feedback during robotic surgery is critical in order to minimize potential injury to the patient and decrease recovery time from surgical procedures. Here we describe the use of a novel strain gauge printing method to apply low profile, low cost sensors directly to the surface of da Vinci surgical robot end effectors (Intuitive Surgical, Inc.) to sense deflection and provide force feedback. This additive, vapor-deposition-based sensor fabrication method is used to deposit strain gauges directly onto the surfaces of the end effectors with minimal disruption to the device and without the need for adhesives or machining operations. Initial experiments characterize sensor performance and indicate the applicability of the proposed approach for force feedback during minimally invasive procedures.

This paper reports a method and preliminary evaluation of a novel null-space least-squares parameter identification method for a fully nonlinear second -order 6-degree-of-freedom (DOF) dynamic process model of an underactuated underwater vehicle (UV) for which both the model parameters and the control-input parameters are unknown. This paper further reports the application of the identified plant models in combined underwater communication and navigation (cooperative navigation) of UVs. We report an approach to model identification that simultaneously identifies 6-DOF UV nonlinear plant-model parameters, control-surface parameters, and thruster-model parameters. We believe this approach is suitable for identifying plant model parameters from data obtained in full-scale experimental trials of UVs in controlled motion. The reported approach to nonlinear model identification of UVs is evaluated in simulation studies. The resulting identified UV plant models are further evaluated in simulated cooperative navigation missions of the UV that are representative of high-precision survey missions. To the best of our knowledge, this paper reports the first method to identify 6-DOF UV model parameters, control-surface parameters, and thruster-model parameters simultaneously.

Human-Robot Collaboration (HRC) aims to develop robots that provide assistance to human workers while performing physical tasks. Such assistance comes in the form of supportive behaviors that are different from the actions part of the task, and that are meant to help a human worker more effectively accomplish the task. Learning how to provide useful behaviors that are tailored to a human peer represents a difficult challenge. This is due to the need of large amounts of training data in the form of real world observations that include information about such preferences. This data needs to encode not only the structure and progression of the task, but also the different workers' preferences with respect to when and what assistance the robot should provide. Our work separates the challenge of learning a model of the task (which requires a large amount of training data) from that of learning supportive behavior preferences for the interaction (which has obvious restrictions for the number of user-provided demonstrations to which we have access). We first learn a hidden Markov model (HMM) from a training set consisting of observed human workers performing the considered task in simulation. We then use this model to predict, while observing the human peer, what supportive behaviors a robot should offer throughout the task. Building upon the hidden state representation, our system is able to learn the supportive behaviors based on as few as five user-annotated demonstrations, learning a personalized supportive behavior model. We evaluate our system on a user study with 14 participants, and show results on par with human-level prediction for the task.

Human-robot interaction strongly benefits from fast, predictive action recognition. For us this is relatively easy but difficult for a robot. To address this problem, here we present a novel prediction algorithm for manipulation action classes in video sequences. Manipulations are first represented using the Enriched Semantic Event Chain (ESEC) framework. This creates a temporal sequence of static and dynamic spatial relations between the objects that take part in the manipulation by which an action can be quickly recognized. We measured performance on 32 ideal as well as real manipulations and compared our method also against a state of the art trajectory-based HMM method for action recognition. We observe that manipulations can be correctly predicted after only (on average) 45% of action's total time and that we are almost twice as fast as the HMM-based method. Finally, we demonstrate the advantage of this framework in a simple robot demonstration comparing two different approaches.

For a robot to manipulate an object, it has to understand the functions and the actions that can be subjected to the object. This set of information is known as affordance of the object. Affordances are generally defined by the geometrical structures and physical properties of the objects. In this paper, we present an affordance detection network (ADNet) for detecting object affordances using multimodal input i.e., RGB-D data. The method is based on the state-of-the-art fully convolutional network with two encoding streams and one decoding stream. In the presented formulation, the network learns powerful discriminative features independently from the RGB and depth images, which enables it to abstract rich photometrical and geometrical properties of the objects. The multimodal encoding is combined at multiple stages of the network using the late-fusion strategy and used is for predicting the potential affordances of the objects.

In this work we present a novel framework that uses deep learning to predict object feature points that are out-of-view in the input image. This system was developed with the application of model-based tracking in mind, particularly in the case of autonomous inspection robots, where only partial views of the object are available. Out-of-view prediction is enabled by applying scaling to the feature point labels during network training. This is combined with a recurrent neural network architecture designed to provide the final prediction layers with rich feature information from across the spatial extent of the input image. To show the versatility of these out-of-view predictions, we describe how to integrate them in both a particle filter tracker and an optimisation based tracker. To evaluate our work we compared our framework with one that predicts only points inside the image. We show that as the amount of the object in view decreases, being able to predict outside the image bounds adds robustness to the final pose estimation.

Robust online incremental SLAM applications require metrics to evaluate the impact of current measurements. Despite its prevalence in graph pruning, information-theoretic metrics solely are insufficient to detect outliers. The optimal value of the objective function is a better choice to detect outliers but cannot be computed unless the problem is solved. In this paper, we show how the objective function change can be predicted in an incremental pose-graph optimization scheme, without actually solving the problem. The predicted objective function change can be used to guide online decisions or detect outliers. Experiments validate the accuracy of the predicted objective function, and an application to outlier detection is also provided, showing its advantages over M-estimators.

We developed a deadbeat foot placement hopping controller for an untethered monopedal robot, Salto-1P. The controller uses a third order Taylor series approximation to an offline dynamic model and performs well on the physical platform. The robot demonstrated precise foot placement even on trajectories with aggressive changes in speed, direction, and height: in a random walk, its error standard deviation was 0.10 m. We establish how foot placement precision is tightly limited by attitude control accuracy, requiring attitude error less than 0.7 degrees for some tasks. We also show how foot placement precision degrades linearly as hopping height increases. These precision results apply to the large class of controllers that prescribe touchdown angle to control running velocity.

The future of automated driving in urban areas will most probably rely on highly accurate road maps. However, the necessary precision of a localization in such maps has so far only been reached using extra, sensor specific feature layers for localization. In this paper we want to show that it is possible to achieve sufficient accuracy without a separate localization layer. Instead, elements are used that are already contained in high-resolution road maps, such as markings and road borders. For this, we introduce a modular approach in which detections from different detection algorithms are associated with elements in the map and then fused to an absolute pose using an Unscented Kalman Filter. We evaluate our approach using a sensor setup that employs a stereo camera, vehicle odometry and a low-cost GNSS module on a 5km test route covering both narrow urban roads and multi-lane main roads under varying weather conditions. The results show that this approach is capable to be used for highly automated driving, showing an accuracy of 0.08m in typical road scenarios and a is available 98% of the time.

Interest in robotic devices for rehabilitation has increased in the last years, due to the increasing number of patients that require rehabilitation therapies, and the need to optimize existing resources. The UHP rehabilitation robot is a multifunctional device that allows to execute robotized therapies for the upper-limb using a simple pantograph based reconfigurable structure and the implementation of advanced position/force control approaches. However, in applications such as rehabilitation, where the robotic device interacts directly with the user, complying with the demands of the users is as important as complying with the functional requirements. Otherwise, the patient will reject the robotic device. Therefore, in this work the pre-clinical validation of the UHP upper-limb rehabilitation robotic platform is presented. 25 subjects of different physical characteristics have participated in the evaluation of the device, evaluating not only the correct behaviour of the device, but also its safety and adaptativity. Results show the correct behaviour of the platform, and a good acceptance rate of the device.

This paper presents the positioning, navigation and awareness (PNA) system developed for the Underwater Robotic Mining System of the !VAMOS! project [1]. It describes the main components of the !VAMOS! system, the PNA sensors in each of those components, the global architecture of the PNA system, and its main subsystems: Position and Navigation, Realtime Mine Modeling, 3D Virtual reality HMI and Real-time grade system. General results and lessons learn during the first mining field trial in Lee Moor, Devon, UK during the months of September and October 2017 are presented.

We present a new time-integrator for modeling the frictional dynamics of articulated bodies. Our formulation represents the configuration of the articulated body using position variables and then uses those variables to model the friction forces between the articulated body and the environment. Our approach corresponds to a Newton-type optimization scheme that is guaranteed to converge so that it is stable with large timestep sizes. We evaluate the accuracy and stability of our time-integrator by comparing it with a conventional formulations based on the Newton-Euler equation and demonstrate the benefits on standard controller-optimization applications. We achieve 3-5 times speedup over a Newton-Euler-based simulator on a CPU. Our approach can be easily parallelized on a GPU and results in additional 4-15 times performance improvement.

Reliable long-term localization is key for robotic systems in dynamic environments. In this paper, we propose a novel approach for long-term localization using 3D LiDARs, coined PoseMap. In essence, we extract distinctive features from range measurements and bundle these into local views along with observation poses. The sensor's trajectory is then estimated in a sliding window fashion by matching current and old features and minimizing the distances in-between. The map representation facilitates finding a suitable set of old features, by selecting the closest local map(s) for matching. Similarly to a visibility analysis, this procedure provides a suitable set of features for localization but at a fraction of the computational cost. PoseMap also allows for updates and extensions of the map at any time by replacing and adding local maps when necessary. We evaluate our approach using two platforms both equipped with a 3D LiDAR and an IMU, demonstrating localization at 8 Hz and robustness to changes in the environment such as moving vehicles and changing vegetation. PoseMap was implemented on an autonomous vehicle allowing it to drive autonomously over a period of 18 months through a mix of industrial and unstructured off-road environments, covering more than 100 kms without a single localization failure.

Pose estimation is a widely explored problem, enabling many robotic tasks such as grasping and manipulation. In this paper, we tackle the problem of pose estimation for objects that exhibit rotational symmetry, which are common in man-made and industrial environments. In particular, our aim is to infer poses for objects not seen at training time, but for which their 3D CAD models are available at test time. Previous work has tackled this problem by learning to compare captured views of real objects with the rendered views of their 3D CAD models, by embedding them in a joint latent space using neural networks. We show that sidestepping the issue of symmetry in this scenario during training leads to poor performance at test time. We propose a model that reasons about rotational symmetry during training by having access to only a small set of symmetry-labeled objects, whereby exploiting a large collection of unlabeled CAD models. We demonstrate that our approach significantly outperforms a naively trained neural network on a new pose dataset containing images of tools and hardware.