ICRA 2020 论文列表

Motion planning for vehicles under the influence of flow fields can benefit from the idea of streamline-based planning, which exploits ideas from fluid dynamics to achieve computational efficiency. Important to such planners is an efficient means of computing the travel distance and direction between two points in free space, but this is difficult to achieve in strong incompressible flows such as ocean currents. We propose two useful distance functions in analytical form that combine Euclidean distance with values of the stream function associated with a flow field, and with an estimation of the strength of the opposing flow between two points. Further, we propose steering heuristics that are useful for steering towards a sampled point. We evaluate these ideas by integrating them with RRT* and comparing the algorithm's performance with state-of-the-art methods in an artificial flow field and in actual ocean prediction data in the region of the dominant East Australian Current between Sydney and Brisbane. Results demonstrate the method's computational efficiency and ability to find high-quality paths outperforming state-of-the-art methods, and show promise for practical use with autonomous marine robots.

Our goal is to investigate whether discrete state space algorithms are a viable solution to continuous alternatives for mapless navigation. To this end we present an approach based on Double Deep Q-Network and employ parallel asynchronous training and a multi-batch Priority Experience Replay to reduce the training time. Experiments show that our method trains faster and outperforms both the continuous Deep Deterministic Policy Gradient and Proximal Policy Optimization algorithms. Moreover, we train the models in a custom environment built on the recent Unity learning toolkit and show that they can be exported on the TurtleBot3 simulator and to the real robot without further training. Overall our optimized method is 40% faster compared to the original discrete algorithm. This setting significantly reduces the training times with respect to the continuous algorithms, maintaining a similar level of success rate hence being a viable alternative for mapless navigation.

Dual-arm robots can overcome grasping force and payload limitations of a single arm by jointly grasping an object. However, if the distribution of mass of the grasped object is not even, each arm will experience different wrenches that can exceed its payload limits. In this work, we consider the problem of balancing the wrenches experienced by a dual-arm robot grasping a rigid tray. The distribution of wrenches among the robot arms changes due to objects being placed on the tray. We present an approach to reduce the wrench imbalance among arms through discrete bimanual manipulation. Our approach is based on sequential sliding motions of the grasp points on the surface of the object, to attain a more balanced configuration. We validate our modeling approach and system design through a set of robot experiments.

Visual dirt detection is becoming an important capability of modern professional cleaning robots both for optimizing their wet cleaning results and for facilitating demand-oriented daily vacuum cleaning. This paper presents a robust, fast, and reliable dirt and office item detection system for these tasks based on an adapted YOLOv3 framework. Its superiority over state-of-the-art dirt detection systems is demonstrated in several experiments. The paper furthermore features a dataset generator for creating any number of realistic training images from a small set of real scene, dirt, and object examples.

The directional mechanical impedance of the human ankle was identified from subjects in a standing posture with varying levels of muscle activity. The impedance modeled the different torque responses to angle perturbations about different axes of rotation. This work proposed a novel impedance model that incorporated the coupling between multiple degrees of freedom of the ankle and was validated theoretically and experimentally. The reconstructed torque had an average variance accounted above 94% across twelve subjects. In addition, the impedance varied between and within trials and this variation was explained by changes in the ankle states, i.e., the ankle angle, torque, and muscle activities. These results have implications in the design of new prostheses controllers and the understanding of the human ankle function.

Scene understanding from images is a challenging problem encountered in autonomous driving. On the object level, while 2D methods have gradually evolved from computing simple bounding boxes to delivering finer grained results like instance segmentations, the 3D family is still dominated by estimating 3D bounding boxes. In this paper, we propose a novel approach to jointly infer the 3D rigid-body poses and shapes of vehicles from a stereo image pair using shape priors. Unlike previous works that geometrically align shapes to point clouds from dense stereo reconstruction, our approach works directly on images by combining a photometric and a silhouette alignment term in the energy function. An adaptive sparse point selection scheme is proposed to efficiently measure the consistency with both terms. In experiments, we show superior performance of our method on 3D pose and shape estimation over the previous geometric approach and demonstrate that our method can also be applied as a refinement step and significantly boost the performances of several state-of-the-art deep learning based 3D object detectors. All related materials and demonstration videos are available at the project page https://vision.in.tum.de/research/vslam/direct-shape.

Fixed-wing unmanned aerial vehicles (UAVs) offer significant performance advantages over rotary-wing UAVs in terms of speed, endurance, and efficiency. However, these vehicles have traditionally been severely limited with regards to maneuverability. In this paper, we present a nonlinear control approach for enabling aerobatic fixed-wing UAVs to maneuver in constrained spaces. Our approach utilizes full-state direct trajectory optimization and a minimalistic, but representative, nonlinear aircraft model to plan aggressive fixed-wing trajectories in real-time at 5 Hz across high angles-of-attack. Randomized motion planning is used to avoid local minima and local-linear feedback is used to compensate for model inaccuracies between updates. We demonstrate our method in hardware and show that both local-linear feedback and re-planning are necessary for successful navigation of a complex environment in the presence of model uncertainty.

In this paper we propose to control a quadrotor through direct acceleration feedback. The proposed method, while simple in form, alleviates the need for accurate estimation of platform parameters such as mass and propeller effectiveness. In order to use efficaciously the noisy acceleration measurements in direct feedback, we propose a novel regression-based filter that exploits the knowledge on the commanded propeller speeds, and extracts smooth platform acceleration with minimal delay. Our tests show that the controller exhibits a few millimeter error when performing real world tasks with fast changing mass and effectiveness, e.g., in pick and place operation and in turbulent conditions. Finally, we benchmark the direct acceleration controller against the PID strategy and show the clear advantage of using high-frequency and low-latency acceleration measurements directly in the control feedback, especially in the case of low frequency position measurements that are typical for real outdoor conditions.

In this work, we propose Dilated Point Convolutions (DPC). In a thorough ablation study, we show that the receptive field size is directly related to the performance of 3D point cloud processing tasks, including semantic segmentation and object classification. Point convolutions are widely used to efficiently process 3D data representations such as point clouds or graphs. However, we observe that the receptive field size of recent point convolutional networks is inherently limited. Our dilated point convolutions alleviate this issue, they significantly increase the receptive field size of point convolutions. Importantly, our dilation mechanism can easily be integrated into most existing point convolutional networks. To evaluate the resulting network architectures, we visualize the receptive field and report competitive scores on popular point cloud benchmarks.

Mapping and localization, preferably from a small number of observations, are fundamental tasks in robotics. We address these tasks by combining spatial structure (differentiable mapping) and end-to-end learning in a novel neural network architecture: the Differentiable Mapping Network (DMN). The DMN constructs a spatially structured view-embedding map and uses it for subsequent visual localization with a particle filter. Since the DMN architecture is end-to-end differentiable, we can jointly learn the map representation and localization using gradient descent. We apply the DMN to sparse visual localization, where a robot needs to localize in a new environment with respect to a small number of images from known viewpoints. We evaluate the DMN using simulated environments and a challenging real-world Street View dataset. We find that the DMN learns effective map representations for visual localization. The benefit of spatial structure increases with larger environments, more viewpoints for mapping, and when training data is scarce. Project website: https://sites.google.com/view/differentiable-mapping.

Modern trajectory optimization based approaches to motion planning are fast, easy to implement, and effective on a wide range of robotics tasks. However, trajectory optimization algorithms have parameters that are typically set in advance (and rarely discussed in detail). Setting these parameters properly can have a significant impact on the practical performance of the algorithm, sometimes making the difference between finding a feasible plan or failing at the task entirely. We propose a method for leveraging past experience to learn how to automatically adapt the parameters of Gaussian Process Motion Planning (GPMP) algorithms. Specifically, we propose a differentiable extension to the GPMP2 algorithm, so that it can be trained end-to-end from data. We perform several experiments that validate our algorithm and illustrate the benefits of our proposed learning-based approach to motion planning.

Consumer demand for augmented reality (AR) in mobile phone applications, such as the Apple ARKit. Such applications have potential to expand access to robot grasp planning systems such as Dex-Net. AR apps use structure from motion methods to compute a point cloud from a sequence of RGB images taken by the camera as it is moved around an object. However, the resulting point clouds are often noisy due to estimation errors. We present a distributed pipeline, Dex-Net AR, that allows point clouds to be uploaded to a server in our lab, cleaned, and evaluated by Dex-Net grasp planner to generate a grasp axis that is returned and displayed as an overlay on the object. We implement Dex-Net AR using the iPhone and ARKit and compare results with those generated with high-performance depth sensors. The success rates with AR on harder adversarial objects are higher than traditional depth images. The server URL is https://sites.google.com/berkeley.edu/dex-net-ar/home.

This paper presents a wheeled wall-climbing robot with a shape-adaptive magnetic adhesion mechanism for large steel structures. To travel up and down various curved ferromagnetic surfaces, we developed a 2 DOF rotational magnetic adhesion mechanism installed on each wheel that can change the orientation of the magnets to keep the magnetic force direction always normal to the contact surface. These magnetic wheels have a spherical shape and can move relative to the main body by a non-elastic suspension mechanism so that the robot can climb up small obstacles on the ground and find contact points for each wheel on a wall with an arbitrary curved shape. Being geometrically stable is important for the robot because this robot is intended to be a mobile base for a welding manipulator. The detailed design of the mechanism and the results of climbing tests are presented.

This paper presents a study on a compliant cable-driven exoskeleton for hip assistance in lifting tasks that is aimed at preventing low-back pain and injuries in the vocational setting. In the proposed concept, we used twisted string actuator (TSA) to design a light-weight and powerful exoskeleton that benefits from inherent TSA advantages. We have noted that nonlinear nature of twisted strings’ transmission ratio (decreasing with twisting) closely matched typical torque-speed requirements for hip assistance during lifting tasks and tried to use this fact in the exoskeleton design and motor selection. Hip-joint torque and speed required to lift a 10-kg load from stoop to stand were calculated, which gave us a baseline that we used to design and manufacture a practical exoskeleton prototype. Preliminary experimental trials demonstrated that the proposed device was capable of generating required torque and speed at the hip joint while weighing under 6 kg, including battery.

With the rapid rise of 3D-printing as a competitive mass manufacturing method, manual "decaking" – i.e. removing the residual powder that sticks to a 3D-printed part – has become a significant bottleneck. Here, we introduce, for the first time to our knowledge, a robotic system for automated decaking of 3D-printed parts. Combining Deep Learning for 3D perception, smart mechanical design, motion planning, and force control for industrial robots, we developed a system that can automatically decake parts in a fast and efficient way. Through a series of decaking experiments performed on parts printed by a Multi Jet Fusion printer, we demonstrated the feasibility of robotic decaking for 3D-printing-based mass manufacturing.

A wheeled robot operating on various terrains such as scattered obstacles and slopes is required to cope with and overcome the driving environment. In this paper, in order to overcome a step-type obstacle and to steadily ascend on the slope, the main body rotation mechanism, which controls the load distribution on the robot wheels was proposed for a wheel-drive robot. By rotating the center of the body mass, the friction/traction force required for climbing step obstacles can be reduced. In the case of slope traveling, the slip can be suppressed, and the traveling ability improved by controlling the load distribution excessively increased on the downhill wheel due to the attitude change of the robot's body. The mechanical effect of the proposed body rotation mechanism was analyzed. In addition, based on the design and manufacture of the robot platform, the effectiveness of the proposed mechanism was convincingly demonstrated by indoor test for step-obstacle climbing and slope-traveling.

Fiducial markers like AprilTags play an important role in robotics, e.g., for the calibration of cameras or the localization of robots. One of the most important properties of an algorithm for detecting such tags is its localization accuracy.In this paper, we present the results of an extensive comparison of four freely available libraries capable of detecting AprilTags, namely AprilTag 3, AprilTags C++, ArUco as standalone libraries, and the OpenCV algorithm based on ArUco. The focus of the comparison is on localization accuracy, but the processing time is also examined. Besides working with pure tags, their extension to checkerboard corners is investigated.In addition, we present two new post-processing techniques. Firstly, a method that can filter out very inaccurate detections resulting from partial border occlusion, and secondly a new highly accurate method for edge refinement. With this we achieve a median pixel error of 0.017 px, compared to 0.17 px for standard OpenCV corner refinement.The dataset used for the evaluation, as well as the developed post-processing techniques, are made publicly available to encourage further comparison and improvement of the detection libraries.

We propose a method for detecting execution anomalies in robotics and autonomy software. The algorithm uses system monitoring techniques to obtain profiles of executions. It uses a clustering algorithm to create clusters of those executions, representing nominal execution. A distance metric determines whether additional execution profiles belong to the existing clusters or should be considered anomalies. The method is suitable for identifying faults in robotics and autonomy systems. We evaluate the technique in simulation on two robotics systems, one of which is a real-world industrial system. We find that our technique works well to detect possibly unsafe behavior in autonomous systems.