ICRA 2018 论文列表

We propose a technique for multi-task learning from demonstration that trains the controller of a low-cost robotic arm to accomplish several complex picking and placing tasks, as well as non-prehensile manipulation. The controller is a recurrent neural network using raw images as input and generating robot arm trajectories, with the parameters shared across the tasks. The controller also combines VAE-GAN-based reconstruction with autoregressive multimodal action prediction. Our results demonstrate that it is possible to learn complex manipulation tasks, such as picking up a towel, wiping an object, and depositing the towel to its previous position, entirely from raw images with direct behavior cloning. We show that weight sharing and reconstruction-based regularization substantially improve generalization and robustness, and training on multiple tasks simultaneously increases the success rate on all tasks.

Service robots are becoming a reality across homes. Still, the range of applications supported by such robots remains tied to their ability to self-localize in the environment. Man-made constructions often have a documented blueprint which can be used as input information for robot localization and smart home applications. However, home environments commonly include movable objects and furniture, which can make localization a complicated task, especially for forward-facing horizontal range-finders. In this paper, we present a new and effective global localization approach for home environments which adapts the notion of free space density to a camera pointing to the ceiling. We exploit the available blueprint information, as well as evidence that ceiling vision can provide robust localization information, even in the presence of occlusions. We perform real-world experiments using a robotic vacuum cleaner equipped with an upward-facing camera in two different apartments across multiple trajectories and compare the proposed method with competing approaches. Our solution shows superior localization results using maps where neither furniture or movable objects are not modeled.

In this paper, we present a collaborative path-planning framework for a group of micro aerial vehicles that are capable of localizing through vision. Each of the micro aerial vehicles is assumed to be equipped with a forward facing monocular camera. The vehicles initially use their captured images to build 3D maps through common features; and subsequently track these features to localize through 3D-2D correspondences. The planning algorithm, while connecting start locations to provided goal locations, also aims to reduce the localization uncertainty of the vehicles in the group. To achieve this, we develop a two-step planning framework: the first step attempts to build an improved map of the environment by solving the next-best-view problem for multiple cameras. We express this as a black-box optimization problem and solve it using the Covariance Matrix Adaption evolution strategy (CMA-ES). Once an improved map is available, the second stage of the planning framework performs belief space planning for the vehicles individually using the rapidly exploring random belief tree (RRBT) algorithm. Through the RRBT approach, the planner generates paths that ensure feature visibility while attempting to optimize path cost and reduce localization uncertainty. We validate our approach using experiments conducted in a high visual-fidelity aerial vehicle simulator, Microsoft AirSim.

The booming interest in Unmanned Aerial Vehicles (UAV s) is fed by their potentially great impact, however progress is hindered by their limited perception capabilities. While vision-based odometry was shown to run successfully onboard UAV s, loop-closure detection to correct for drift or to recover from tracking failures, has so far, proven particularly challenging for UAVs. At the heart of this is the problem of viewpoint-tolerant place recognition; in stark difference to ground robots, UAVs can revisit a scene from very different viewpoints. As a result, existing approaches struggle greatly as the task at hand violates underlying assumptions in assessing scene similarity. In this paper, we propose a place recognition framework, which exploits both efficient binary features and noisy estimates of the local 3D geometry, which are anyway computed for visual-inertial odometry onboard the UAV. Attaching both an appearance and a geometry signature to each `location', the proposed approach demonstrates unprecedented recall for perfect precision as well as high quality loop-closing transformations on both flying and hand-held datasets exhibiting large viewpoint and appearance changes as well as perceptual aliasing.

Vision and touch are two of the important sensing modalities for humans and they offer complementary information for sensing the environment. Robots could also benefit from such multi-modal sensing ability. In this paper, addressing for the first time (to the best of our knowledge) texture recognition from tactile images and vision, we propose a new fusion method named Deep Maximum Covariance Analysis (DMCA) to learn a joint latent space for sharing features through vision and tactile sensing. The features of camera images and tactile data acquired from a GelSight sensor are learned by deep neural networks. But the learned features are of a high dimensionality and are redundant due to the differences between the two sensing modalities, which deteriorates the perception performance. To address this, the learned features are paired using maximum covariance analysis. Results of the algorithm on a newly collected dataset of paired visual and tactile data relating to cloth textures show that a good recognition performance of greater than 90% can be achieved by using the proposed DMCA framework. In addition, we find that the perception performance of either vision or tactile sensing can be improved by employing the shared representation space, compared to learning from unimodal data.

Recent successes in reinforcement learning have lead to the development of complex controllers for realworld robots. As these robots are deployed in safety-critical applications and interact with humans, it becomes critical to ensure safety in order to avoid causing harm. A first step in this direction is to test the controllers in simulation. To be able to do this, we need to capture what we mean by safety and then efficiently search the space of all behaviors to see if they are safe. In this paper, we present an active-testing framework based on Bayesian Optimization. We specify safety constraints using logic and exploit structure in the problem in order to test the system for adversarial counter examples that violate the safety specifications. These specifications are defined as complex boolean combinations of smooth functions on the trajectories and, unlike reward functions in reinforcement learning, are expressive and impose hard constraints on the system. In our framework, we exploit regularity assumptions on individual functions in form of a Gaussian Process (GP) prior. We combine these into a coherent optimization framework using problem structure. The resulting algorithm is able to provably verify complex safety specifications or alternatively find counter examples. Experimental results show that the proposed method is able to find adversarial examples quickly.

We present a real-time vehicle detection and tracking system to accomplish the complex task of driving behavior analysis in urban environments. We propose a robust fusion system that combines a monocular camera and a 2D Lidar. This system takes advantage of three key components: robust vehicle detection using deep learning techniques, high precision range estimation from Lidar, and road context from the prior map knowledge. The camera and Lidar sensor fusion, data association and track management are all performed in the global map coordinate system by taking into account the sensors' characteristics. Lastly, behavior reasoning is performed by examining the tracked vehicle states in the lane coordinate system in which the road context is encoded. We validated our approach by tracking a leading vehicle while it performed usual urban driving behaviors such as lane keeping, stop-and-go at intersections, lane changing, overtaking and turning. The leading vehicle was tracked consistently throughout the 2.3 km route and its behavior was classified reliably.

In this paper, we introduce a novel robotic prosthetic knee as shown in Fig. 1 (named as SuKnee) with variable transmission mechanism that could vary transmission ratio while knee angle varies during ambulation activities. A slider crank mechanism is utilized to transform linear motion of series elastic actuator to rotary motion of knee joint. And it contributes to variable transmission ratio with knee angle, which help obtain desired speed variation and torque output in different activities in one mechanism. This feature could uniquely give the SuKnee both: the torque necessary to assist with standing up from a chair and the speed necessary to swing the leg forward during walking. The knee has an active mode, where it operates with batteries and is capable of providing external power, and a passive mode, behaving like a passive prosthesis. Preliminary tests have been performed by a transfemoral amputee and SuKnee could provide user with power to assist walking on level ground and standing up from a chair. And a passive mode test shows it could work like passive prosthesis after battery exhaustion.

Robots mounted on an unmanned chaser satellite could be used for performing rendezvous and grasping manoeuvres in order to repair satellites or remove space debris from orbit. Use of manipulators for such purposes is challenging, since the performed task need to be done autonomously, accurately and with high level of robustness. During manoeuvres high disturbances might appear e.g. due to contact between the manipulator arm end-effector and the target spacecraft. In this paper an approach for validation of the robotic subsystem during chaser rendezvous and grasping manoeuvre has been shown. Two type of testbed systems were used: planar air-bearing microgravity simulators and a test-bed system with industrial robots. The proposed approach took advantage of possible replacement of particular subsystem in reference model or reference hardware in specified test-bed system. The list of tests performed are included in the paper.

This paper presents a simple and robust method for the automatic localisation of static 3D objects in large-scale urban environments. By exploiting the potential to merge a large volume of noisy but accurately localised 2D image data, we achieve superior performance in terms of both robustness and accuracy of the recovered 3D information. The method is based on a simple distributed voting schema which can be fully distributed and parallelised to scale to large-scale scenarios. To evaluate the method we collected city-scale data sets from New York City and San Francisco consisting of almost 400k images spanning the area of 40 km2 and used it to accurately recover the 3D positions of traffic lights. We demonstrate a robust performance and also show that the solution improves in quality over time as the amount of data increases.

Predictive control is an efficient model-based methodology to control complex dynamical systems. In general, it boils down to the resolution at each control cycle of a large nonlinear optimization problem. A critical issue is then to provide a good guess to initialize the nonlinear solver so as to speed up convergence. This is particularly important when disturbances or changes in the environment prevent the use of the trajectory computed at the previous control cycle as initial guess. In this paper, we introduce an original and very efficient solution to automatically build this initial guess. We propose to rely on off-line computation to build an approximation of the optimal trajectories, that can be used on-line to initialize the predictive controller. To that end, we combined the use of sampling-based planning, policy learning with generic representations (such as neural networks), and direct optimal control. We first propose an algorithm to simultaneously build a kinodynamic probabilistic roadmap (PRM) and approximate value function and control policy. This algorithm quickly converges toward an approximation of the optimal state-control trajectories (along with an optimal PRM). Then, we propose two methods to store the optimal trajectories and use them to initialize the predictive controller. We experimentally show that directly storing the state-control trajectories leads the predictive controller to quickly converges (2 to 5 iterations) toward the (global) optimal solution. The results are validated in simulation with an unmanned aerial vehicle (UAV) and other dynamical systems.

Instrumenting and collecting annotated visual grasping datasets to train modern machine learning algorithms can be extremely time-consuming and expensive. An appealing alternative is to use off-the-shelf simulators to render synthetic data for which ground-truth annotations are generated automatically. Unfortunately, models trained purely on simulated data often fail to generalize to the real world. We study how randomized simulated environments and domain adaptation methods can be extended to train a grasping system to grasp novel objects from raw monocular RGB images. We extensively evaluate our approaches with a total of more than 25,000 physical test grasps, studying a range of simulation conditions and domain adaptation methods, including a novel extension of pixel-level domain adaptation that we term the GraspGAN. We show that, by using synthetic data and domain adaptation, we are able to reduce the number of real-world samples needed to achieve a given level of performance by up to 50 times, using only randomly generated simulated objects. We also show that by using only unlabeled real-world data and our GraspGAN methodology, we obtain real-world grasping performance without any real-world labels that is similar to that achieved with 939,777 labeled real-world samples.

The most data-efficient algorithms for reinforcement learning in robotics are model-based policy search algorithms, which alternate between learning a dynamical model of the robot and optimizing a policy to maximize the expected return given the model and its uncertainties. Among the few proposed approaches, the recently introduced Black-DROPS algorithm exploits a black-box optimization algorithm to achieve both high data-efficiency and good computation times when several cores are used; nevertheless, like all model-based policy search approaches, Black-DROPS does not scale to high dimensional state/action spaces. In this paper, we introduce a new model learning procedure in Black-DROPS that leverages parameterized black-box priors to (1) scale up to high-dimensional systems, and (2) be robust to large inaccuracies of the prior information. We demonstrate the effectiveness of our approach with the “pendubot” swing-up task in simulation and with a physical hexapod robot (48D state space, 18D action space) that has to walk forward as fast as possible. The results show that our new algorithm is more data-efficient than previous model-based policy search algorithms (with and without priors) and that it can allow a physical 6-legged robot to learn new gaits in only 16 to 30 seconds of interaction time.

Robotic swarms are decentralized multi-robot systems whose members use local information from proximal neighbors to execute simple reactive control laws that result in emergent collective behaviors. In contrast, members of a general multi-robot system may have access to global information, all-to-all communication or sophisticated deliberative collaboration. Some algorithms in the literature are applicable to robotic swarms. Others require the extra complexity of general multi-robot systems. Given an application domain, a system designer or supervisory operator must choose an appropriate system or algorithm respectively that will enable them to achieve their goals while satisfying mission constraints (e.g, bandwidth, energy, time limits). In this paper, we compare representative swarm and general multi-robot algorithms in two application domains - navigation and dynamic area coverage - with respect to several metrics (e.g, completion time, distance travelled). Our objective is to characterize each class of algorithms to inform offline system design decisions by engineers or online algorithm selection decisions by supervisory operators. Our contributions are (a) an empirical performance comparison of representative swarm and general multi-robot algorithms in two application domains, (b) a comparative analysis of the algorithms based on the theory of information invariants, which provides a theoretical characterization supported by our emnirical results.

Specialization has always been a tool for work distribution and simplification in nature and in distributed robotics. We present a novel approach to use hardware specialization hierarchically to enhance the capabilities of a swarm without increasing complexity, allowing a numerous group of robots to benefit from the extended features of a few to complete a task that was impossible for them before. We tested the concept under a simulated environment with a classical distributed robotics problem, shape formation, and validated the simulated results against a real experiment.

Most of the motion re-targeting techniques are grounded on virtual character animation research, which means that they typically assume that the target embodiment has unconstrained joint angular velocities. However, because robots often do have such constraints, traditional re-targeting approaches can originate irregular delays in the robot motion. With the goal of ensuring synchronization between verbal and nonverbal behavior, this paper proposes an optimization framework for processing re-targeted motion sequences that addresses constraints such as joint angle and angular velocities. The proposed framework was evaluated on a humanoid robot using both objective and subjective metrics. While the analysis of the joint motion trajectories provides evidence that our framework successfully performs the desired modifications to ensure verbal and nonverbal behavior synchronization, results from a perceptual study showed that participants found the robot motion generated by our method more natural, elegant and lifelike than a control condition.

Contact forces of the hand are visually unobservable, but play a crucial role in understanding hand-object interactions. In this paper, we propose an unsupervised learning approach for manipulation event segmentation and manipulation event parsing. The proposed framework incorporates hand pose kinematics and contact forces using a low-cost easy-to-replicate tactile glove. We use a temporal grammar model to capture the hierarchical structure of events, integrating extracted force vectors from the raw sensory input of poses and forces. The temporal grammar is represented as a temporal And-Or graph (T-AOG), which can be induced in an unsupervised manner. We obtain the event labeling sequences by measuring the similarity between segments using the Dynamic Time Alignment Kernel (DTAK). Experimental results show that our method achieves high accuracy in manipulation event segmentation, recognition and parsing by utilizing both pose and force data.

This work presents a method for contact state estimation using fuzzy clustering to learn contact probability for full, six-dimensional humanoid contacts. The data required for training is solely from proprioceptive sensors - endeffector contact wrench sensors and inertial measurement units (IMUs) - and the method is completely unsupervised. The resulting cluster means are used to efficiently compute the probability of contact in each of the six endeffector degrees of freedom (DoFs) independently. This clustering-based contact probability estimator is validated in a kinematics-based base state estimator in a simulation environment with realistic added sensor noise for locomotion over rough, low-friction terrain on which the robot is subject to foot slip and rotation. The proposed base state estimator which utilizes these six DoF contact probability estimates is shown to perform considerably better than that which determines kinematic contact constraints purely based on measured normal force.

Hand synergies, or joint coordination patterns, have become an effective tool for achieving versatile robotic grasping with simple hands or planning algorithms. Here we propose a method to determine the hand synergies such that they can be physically implemented in an underactuated fashion. Given a kinematic hand model and a set of desired grasps, our algorithm optimizes a Mechanically Realizable Manifold designed to be achievable by a physical underactuation mechanism, enabling the resulting hand to achieve the desired grasps with few actuators. Furthermore, in contrast to existing methods for determining synergies which are only concerned with hand posture, our method explicitly optimizes the stability of the target grasps. We implement this method in the design of a three-finger single-actuator hand as an example, and evaluate its effectiveness numerically and experimentally.