ICRA 2015 论文列表

We develop a new paradigm for designing fully streaming, area-efficient FPGA implementations of common building blocks for vision algorithm. By focusing on avoiding redundant computation we achieve a reduction of one to two orders of magnitude reduction in design area utilization as compared to previous implementations. We demonstrate that our design works in practice by building five 325 frames per second, high resolution Harris corner detection cores onto a single FPGA.

In this paper, we present algorithms for self-reconfiguration of modular robots that move by pivoting. The modules are cubes that can pivot about their edges along the x̂, ŷ, or ẑ axes to move on a 3-dimensional substrate. This is a different model from prior work, which usually considers modules that slide along their faces. We analyze the pivoting cube model and give sufficient conditions for reconfiguration to be feasible. In particular, we show that if an initial configuration does not contain any of three subconfigurations, which we call rules, then it can reconfigure into a line. We provide provably correct algorithms for reconfiguration for both 2-D and 3-D systems, and we verify our algorithms via simulation on randomly generated 2-D and 3-D configurations.

This paper presents experimentally realized bipedal robotic walking using ideal torque controllers via a novel approach termed the ideal model resolved motion method (IM-RMM), where a system's ideal closed-loop dynamics are integrated forward from the actual state of the hardware to provide desired positions and velocity commands to a PD controller. By combining this method with gaits generated using the Human-Inspired Control framework, walking was realized experimentally on the DURUS platform, designed and built by SRI, and achieved with minimal system identification. For comparison, two controllers, one using feedback linearization and one using Control Lyapunov Function based Quadratic Programs (CLF-QP), both realized through IM-RMM, are compared with a benchmark procedure, the Hybrid Zero Dynamics reconstruction, that is shown to provide reliable walking in literature. The results of both simulations and experiments are presented, with the CLF-QP implemented via IM-RMM resulting in the lowest experimental specific energetic cost of transport of cet = 0.63 achieved during sustained walking on the 31.5 kg bipedal robot.

In this paper, we present a novel approach for tracking a deformable anatomical target within 3D ultrasound volumes. Our method is able to estimate deformations caused by the physiological motions of the patient. The displacements of moving structures are estimated from an intensity-based approach combined with a physically-based model and has therefore the advantage to be less sensitive to the image noise. Furthermore, our method does not use any fiducial marker and has real-time capabilities. The accuracy of our method is evaluated on real data acquired from an organic phantom. The validation is performed on different types of motions comprising rigid and non-rigid motions. Thus, our approach opens novel possibilities for computer-assisted interventions where deformable organs are involved.

Real-time and reliable localization is a prerequisite for autonomously performing high-level tasks with micro aerial vehicles(MAVs). Nowadays, most existing methods use vision system for 6DoF pose estimation, which can not work in degraded visual environments. This paper presents an onboard 6DoF pose estimation method for an indoor MAV in challenging GPS-denied degraded visual environments by using a RGB-D camera. In our system, depth images are mainly used for odometry estimation and localization. First, a fast and robust relative pose estimation (6DoF Odometry) method is proposed, which uses the range rate constraint equation and photometric error metric to get the frame-to-frame transform. Then, an absolute pose estimation (6DoF Localization) method is proposed to locate the MAV in a given 3D global map by using a particle filter. The whole localization system can run in real-time on an embedded computer with low CPU usage. We demonstrate the effectiveness of our system in extensive real environments on a customized MAV platform. The experimental results show that our localization system can robustly and accurately locate the robot in various practical challenging environments.

This paper proposes a method for real-time de-blurring of images captured by a robotic system. De-blurring is achieved by combining a dynamics-based approach and parallel computing. Existing methods have successfully restored a blurry image but were not suitable for real-time applications. The dynamics-based method with parallel programming in the FPGA is used to estimate the PSF during the camera exposure window. The deconvolution process, which involves iterative matrix calculations with millions of pixels, is then performed on the GPU to achieve real-time performance. The proposed method has been evaluated by using a robotic system which has a total of 32 piezoelectric actuators and one vision sensor. The frame rate of image de-blurring ranges from 20.32 to 40.68 fps and thus real-time performance is achieved. In comparison to existing methods, the proposed method produces the best image quality. The dynamics-based approach is extended to real-time image stitching which is computationally cheap and robust to motion blur. It is demonstrated from image stitching that real-time de-blurring is beneficial for image processing.

We present an accurate, real-time approach to robotic grasp detection based on convolutional neural networks. Our network performs single-stage regression to graspable bounding boxes without using standard sliding window or region proposal techniques. The model outperforms state-of-the-art approaches by 14 percentage points and runs at 13 frames per second on a GPU. Our network can simultaneously perform classification so that in a single step it recognizes the object and finds a good grasp rectangle. A modification to this model predicts multiple grasps per object by using a locally constrained prediction mechanism. The locally constrained model performs significantly better, especially on objects that can be grasped in a variety of ways.

Understanding social context is an important skill for robots that share a space with humans. In this paper, we address the problem of recognizing gender, a key piece of information when interacting with people and understanding human social relations and rules. Unlike previous work which typically considered faces or frontal body views in image data, we address the problem of recognizing gender in RGB-D data from side and back views as well. We present a large, gender-balanced, annotated, multi-perspective RGB-D dataset with full-body views of over a hundred different persons captured with both the Kinect v1 and Kinect v2 sensor. We then learn and compare several classifiers on the Kinect v2 data using a HOG baseline, two state-of-the-art deep-learning methods, and a recent tessellation-based learning approach. Originally developed for person detection in 3D data, the latter is able to learn the best selection, location and scale of a set of simple point cloud features. We show that for gender recognition, it outperforms the other approaches for both standing and walking people while being very efficient to compute with classification rates up to 150 Hz.

We consider the problem of discovering and representing the topology of a modular self-reconfigurable robot (MSR) in which modules do not have a priori information about other modules that belong to the same configuration. We propose a solution that combines two main features of modules - their geometric shape and their inter-connection with other modules, represented as a connectivity graph. We describe a distributed algorithm that can be used by each module to construct the connectivity graph by discovering other modules in four steps - sharing IDs, creating a local configuration structure (LCS) or list of local neighbors, sharing a matrix with LCSs to adjacent modules, and building an adjacency matrix by combining LCSs. Data is exchanged between modules using infra-red (IR) communication. Our proposed technique can operate within limited computational resources available on a module and is robust to module failures. We have tested and demonstrated the successful operation of our proposed technique on the ModRED (Modular Robot for Exploration and Discovery) platform and shown that modules are able to discover both linear and branched configurations using our algorithm.

Finite Element analysis can provide accurate deformable models for soft-robots. However, using such models is very difficult in a real-time system of control. In this paper, we introduce a generic solution that enables a high-rate control and that is compatible with strong real-time constraints. From a Finite Element analysis, computed at low rate, an inverse model of the robot outputs the setpoint values for the actuator in order to obtain a desired trajectory. This inverse problem uses a QP (quadratic-programming) algorithm based on the equations set by the Finite Element Method. To improve the update rate performances, we propose an asynchronous simulation framework that provides a better trade-off between the deformation accuracy and the computational burden. Complex computations such as accurate FEM deformations are done at low frequency while the control is performed at high frequency with strong real-time constraints. The two simulation loops (high frequency and low frequency loops) are mechanically coupled in order to guarantee mechanical accuracy of the system over time. Finally, the validity of the multi-rate simulation is discussed based on measurements of the evolution in the QP matrix and an experimental validation is conducted to validate the correctness of the high-rate inverse model on a real robot.

In this work, we propose a constraint-based algorithm for combined trajectory generation and kinematic control for robotic manipulators. The main feature of the algorithm is to ease robot programming, shifting from an imperative paradigm, in which task constraints are semantically and uniquely mapped into a suitable end-effector velocity profile, towards a declarative motion programming, where such constraints are turned by the controller into motion commands only at run-time: The system embeds the capability of handling real-time events, such as updated sensor readings, with reduced pre-programmed control logics. An experimental case study based on a 7-DOF robot demonstrates the effectiveness of the approach.

High speed, low latency obstacle avoidance is essential for enabling Micro Aerial Vehicles (MAVs) to function in cluttered and dynamic environments. While other systems exist that do high-level mapping and 3D path planning for obstacle avoidance, most of these systems require high-powered CPUs on-board or off-board control from a ground station.

Reachable volumes are a new technique that allows one to efficiently restrict sampling to feasible/reachable regions of the planning space even for high degree of freedom and highly constrained problems. However, they have so far only been applied to graph-based sampling-based planners. In this paper we develop the methodology to apply reachable volumes to tree-based planners such as Rapidly-Exploring Random Trees (RRTs). In particular, we propose a reachable volume RRT called RVRRT that can solve high degree of freedom problems and problems with constraints. To do so, we develop a reachable volume stepping function, a reachable volume expand function, and a distance metric based on these operations. We also present a reachable volume local planner to ensure that local paths satisfy constraints for methods such as PRMs. We show experimentally that RVRRTs can solve constrained problems with as many as 64 degrees of freedom and unconstrained problems with as many as 134 degrees of freedom. RVRRTs can solve problems more efficiently than existing methods, requiring fewer nodes and collision detection calls. We also show that it is capable of solving difficult problems that existing methods cannot.

We consider the problem of building high-quality 3D object models from commodity RGB and depth sensors. Applications of such a database include instance and object recognition, robot grasping, virtual reality, graphics, and online shopping. Unfortunately, modern reconstruction approaches have difficulties in reconstructing objects with major transparencies (e.g., KinectFusion [22]) and/or concavities (e.g., visual hull). This paper presents a method to fuse visual hull information from off-the-shelf RGB cameras and KinectFusion cues from commodity depth sensors to produce models that are substantially better than either approach on its own. Extensive experiments on the recently published BigBIRD dataset [25] demonstrate that our reconstructions recover more accurate shape and detail than competing approaches, particularly on challenging objects with transparencies and/or concavities. Quantitative evaluations indicate that our approach consistently outperforms competing methods and achieves under 2 mm RMS error. We plan to release our code after the review process.

In this paper we introduce a new geometric calibration algorithm, and a geometric method of 3D reconstruction using a panoramic microwave radar and a camera. These two sensors are complementary, considering the robustness to environmental conditions and depth detection ability of the radar on one hand, and the high spatial resolution of a vision sensor on the other hand. This makes the approach well adapted for large scale outdoor cartography. Firstly, we address the global calibration problem which consists in finding the exact transformation between radar and camera coordinate systems. The method is based on the optimization of a non-linear criterion obtained from a set of radar-to-image target correspondences. Unlike existing methods, no special configuration of the 3D points is required, only the knowledge of inter-targets distance is needed. This makes the method flexible and easy to use by a non expert operator. Secondly, we present a 3D reconstruction method based on sensors geometry. Both methods have been validated with synthetic and real data.

RPN (Robotic Programming Network) is an initiative to bring existing remote robot laboratories to a new dimension, by adding the flexibility and power of writing ROS code in an Internet browser and running it in the remote robot with a single click. The code is executed in the robot server at full speed, i.e. without any communication delay, and the output of the process is returned back. Built upon Robot Web Tools, RPN works out-of-the-box in any ROS-based robot or simulator. This paper presents the core functionality of RPN in the context of a web-enabled ROS system, its possibilities for remote education and training, and some experimentation with simulators and real robots in which we have integrated the tool in a Moodle environment, creating some programming courses and make it open to researchers and students (http://robotprogramming.uji.es).

We describe a novel RGBD relocalisation algorithm based on key point matching. It combines two components. First, a graph matching algorithm which takes into account the pairwise 3-D geometry amongst the key points, giving robust relocalisation. Second, a point selection process which provides an even distribution of the ‘most matchable’ points across the scene based on non-maximum suppression within voxels of a volumetric grid. This ensures a bounded set of matchable key points which enables tractable and scalable graph matching at frame rate. We present evaluations using a public dataset and our own more difficult dataset containing large pose changes, fast motion and non-stationary objects. It is shown that the method significantly out performs state-of-the-art methods.

Object recognition and pose estimation from RGB-D images are important tasks for manipulation robots which can be learned from examples. Creating and annotating datasets for learning is expensive, however. We address this problem with transfer learning from deep convolutional neural networks (CNN) that are pre-trained for image categorization and provide a rich, semantically meaningful feature set. We incorporate depth information, which the CNN was not trained with, by rendering objects from a canonical perspective and colorizing the depth channel according to distance from the object center. We evaluate our approach on the Washington RGB-D Objects dataset, where we find that the generated feature set naturally separates classes and instances well and retains pose manifolds. We outperform state-of-the-art on a number of subtasks and show that our approach can yield superior results when only little training data is available.

We present a novel, real-time algorithm to extract the trajectory of each pedestrian in moderately dense crowd videos. In order to improve the tracking accuracy, we use a hybrid motion model that combines discrete and continuous flow models. The discrete model is based on microscopic agent formulation and is used for local navigation, interaction, and collision avoidance. The continuum model accounts for macroscopic behaviors, including crowd orientation and flow. We use our hybrid model with particle filters to compute the trajectories at interactive rates. We demonstrate its performance in moderately-dense crowd videos with tens of pedestrians and highlight the improved accuracy on different datasets.