IROS 2020 论文列表

Mistakes/uncertainties in object detection could lead to catastrophes when deploying robots in the real world. In this paper, we measure the uncertainties of object localization to minimize this kind of risk. Uncertainties emerge upon challenging cases like occlusion. The bounding box borders of an occluded object can have multiple plausible configurations. We propose a deep multivariate mixture of Gaussians model for probabilistic object detection. The covariances help to learn the relationship between the borders, and the mixture components potentially learn different configurations of an occluded part. Quantitatively, our model improves the AP of the baselines by 3.9% and 1.4% on CrowdHuman and MS-COCO respectively with almost no computational or memory overhead. Qualitatively, our model enjoys explainability since the resulting covariance matrices and the mixture components help measure uncertainties.

By virtue of their ultra high resolution, scanning electron microscopes (SEMs) are essential to study topography, morphology, composition, and crystallography of materials, and thus are widely used for advanced researches in physics, chemistry, pharmacy, geology, etc. The major hindrance of using SEMs is that obtaining high quality images from SEMs requires a professional control of many control parameters. Therefore, it is not an easy task even for an experienced researcher to get high quality sample images without any help from SEM experts. In this paper, we propose and implement a deep learning-based autonomous SEM machine, which assesses image quality and controls parameters autonomously to get high quality sample images just as if human experts do. This world's first autonomous SEM machine may be the first step to bring SEMs, previously used only for advanced researches due to its difficulty in use, into much broader applications such as education, manufacture, and mechanical diagnosis, which are previously meant for optical microscopes.

Estimating relative camera poses from consecutive frames is a fundamental problem in visual odometry (VO) and simultaneous localization and mapping (SLAM), where classic methods consisting of hand-crafted features and sampling-based outlier rejection have been a dominant choice for over a decade. Although multiple works propose to replace these modules with learning-based counterparts, most have not yet been as accurate, robust and generalizable as conventional methods. In this paper, we design an end-to-end trainable framework consisting of learnable modules for detection, feature extraction, matching and outlier rejection, while directly optimizing for the geometric pose objective. We show both quantitatively and qualitatively that pose estimation performance may be achieved on par with the classic pipeline. Moreover, we are able to show by end-to-end training, the key components of the pipeline could be significantly improved, which leads to better generalizability to unseen datasets compared to existing learning-based methods.

With the recent boost in autonomous driving, increased attention has been paid on radars as an input for occupancy mapping. Besides their many benefits, the inference of occupied space based on radar detections is notoriously difficult because of the data sparsity and the environment dependent noise (e.g. multipath reflections). Recently, deep learning-based inverse sensor models, from here on called deep ISMs, have been shown to improve over their geometric counterparts in retrieving occupancy information [1], [2], [3]. Nevertheless, these methods perform a data-driven interpolation which has to be verified later on in the presence of measurements. In this work, we describe a novel approach to integrate deep ISMs together with geometric ISMs into the evidential occupancy mapping framework. Our method leverages both the capabilities of the data-driven approach to initialize cells not yet observable for the geometric model effectively enhancing the perception field and convergence speed, while at the same time use the precision of the geometric ISM to converge to sharp boundaries. We further define a lower limit on the deep ISM estimate's certainty together with analytical proofs of convergence which we use to distinguish cells that are solely allocated by the deep ISM from cells already verified using the geometric approach.

Sequential pulling policies to flatten and smooth fabrics have applications from surgery to manufacturing to home tasks such as bed making and folding clothes. Due to the complexity of fabric states and dynamics, we apply deep imitation learning to learn policies that, given color (RGB), depth (D), or combined color-depth (RGBD) images of a rectangular fabric sample, estimate pick points and pull vectors to spread the fabric to maximize coverage. To generate data, we develop a fabric simulator and an algorithmic supervisor that has access to complete state information. We train policies in simulation using domain randomization and dataset aggregation (DAgger) on three tiers of difficulty in the initial randomized configuration. We present results comparing five baseline policies to learned policies and report systematic comparisons of RGB vs D vs RGBD images as inputs. In simulation, learned policies achieve comparable or superior performance to analytic baselines. In 180 physical experiments with the da Vinci Research Kit (dVRK) surgical robot, RGBD policies trained in simulation attain coverage of 83% to 95% depending on difficulty tier, suggesting that effective fabric smoothing policies can be learned from an algorithmic supervisor and that depth sensing is a valuable addition to color alone. Supplementary material is available at https://sites.google.com/view/fabric-smoothing.

For in-hand manipulation, estimation of the object pose inside the hand is one of the important functions to manipulate objects to the target pose. Since in-hand manipulation tends to cause occlusions by the hand or the object itself, image information only is not sufficient for in-hand object pose estimation. Multiple modalities can be used in this case, the advantage is that other modalities can compensate for occlusion, noise, and sensor malfunctions. Even though deciding the utilization rate of a modality (referred to as reliability value) corresponding to the situations is important, the manual design of such models is difficult, especially for various situations. In this paper, we propose deep gated multi-modal learning, which self-determines the reliability value of each modality through end-to-end deep learning. For the experiments, an RGB camera and a GelSight tactile sensor were attached to the parallel gripper of the Sawyer robot, and the object pose changes were estimated during grasping. A total of 15 objects were used in the experiments. In the proposed model, the reliability values of the modalities were determined according to the noise level and failure of each modality, and it was confirmed that the pose change was estimated even for unknown objects.

This paper addresses the problem of learning to complete a scene's depth from sparse depth points and images of indoor scenes. Specifically, we study the case in which the sparse depth is computed from a visual-inertial simultaneous localization and mapping (VI-SLAM) system. The resulting point cloud has low density, it is noisy, and has nonuniform spatial distribution, as compared to the input from active depth sensors, e.g., LiDAR or Kinect. Since the VI-SLAM produces point clouds only over textured areas, we compensate for the missing depth of the low-texture surfaces by leveraging their planar structures and their surface normals which is an important intermediate representation. The pre-trained surface normal network, however, suffers from large performance degradation when there is a significant difference in the viewing direction (especially the roll angle) of the test image as compared to the trained ones. To address this limitation, we use the available gravity estimate from the VI-SLAM to warp the input image to the orientation prevailing in the training dataset. This results in a significant performance gain for the surface normal estimate, and thus the dense depth estimates. Finally, we show that our method outperforms other state-of-the-art approaches both on training (ScanNet [1] and NYUv2 [2]) and testing (collected with Azure Kinect [3]) datasets.

Deep learning in combination with improved training techniques and high computational power has led to recent advances in the field of reinforcement learning (RL) and to successful robotic RL applications such as in-hand manipulation. However, most robotic RL relies on a well known initial state distribution. In real-world tasks, this information is however often not available. For example, when disentangling waste objects the actual position of the robot w.r.t. the objects may not match the positions the RL policy was trained for. To solve this problem, we present a novel adversarial reinforcement learning (ARL) framework. The ARL framework utilizes an adversary, which is trained to steer the original agent, the protagonist, to challenging states. We train the protagonist and the adversary jointly to allow them to adapt to the changing policy of their opponent. We show that our method can generalize from training to test scenarios by training an end-to-end system for robot control to solve a challenging object disentangling task. Experiments with a KUKA LBR+ 7-DOF robot arm show that our approach outperforms the baseline method in disentangling when starting from different initial states than provided during training.

We investigate a multi-agent planning problem, where each agent aims to achieve an individual task while avoiding collisions with others. We assume that each agent's task is expressed as a Time-Window Temporal Logic (TWTL) specification defined over a 3D environment. We propose a decentralized receding horizon algorithm for online planning of trajectories. We show that when the environment is sufficiently connected, the resulting agent trajectories are always safe (collision-free) and lead to the satisfaction of the TWTL specifications or their finite temporal relaxations. Accordingly, deadlocks are always avoided and each agent is guaranteed to safely achieve its task with a finite time-delay in the worst case. Performance of the proposed algorithm is demonstrated via numerical simulations and experiments with quadrotors.

Cooperative robotics is a trending topic nowadays as it makes possible a number of tasks that cannot be performed by individual robots, such as heavy payload transportation and agile manipulation. In this work, we address the problem of cooperative transportation by heterogeneous, manipulator- endowed robots. Specifically, we consider a generic number of robotic agents simultaneously grasping an object, which is to be transported to a prescribed set point while avoiding obstacles. The procedure is based on a decentralized leader-follower Model Predictive Control scheme, where a designated leader agent is responsible for generating a trajectory compatible with its dynamics, and the followers must compute a trajectory for their own manipulators that aims at minimizing the internal forces and torques that might be applied to the object by the different grippers. The Model Predictive Control approach appears to be well suited to solve such a problem, because it provides both a control law and a technique to generate trajectories, which can be shared among the agents. The proposed algorithm is implemented using a system comprised of a ground and an aerial robot, both in the robotic Gazebo simulator as well as in experiments with real robots, where the methodological approach is assessed and the controller design is shown to be effective for the cooperative transportation task.

Locomotion is a prime example for adaptive behavior in animals and biological control principles have inspired control architectures for legged robots. While machine learning has been successfully applied to many tasks in recent years, Deep Reinforcement Learning approaches still appear to struggle when applied to real world robots in continuous control tasks and in particular do not appear as robust solutions that can handle uncertainties well. Therefore, there is a new interest in incorporating biological principles into such learning architectures. While inducing a hierarchical organization as found in motor control has shown already some success, we here propose a decentralized organization as found in insect motor control for coordination of different legs. A decentralized and distributed architecture is introduced on a simulated hexapod robot and the details of the controller are learned through Deep Reinforcement Learning. We first show that such a concurrent local structure is able to learn better walking behavior. Secondly, that the simpler organization is learned faster compared to holistic approaches.

Although legged robots are becoming more nonlinear with higher degrees of freedom (DOFs), the centralized nonlinear control methods required to achieve stable locomotion cannot scale with the dimensionality of these robots. This paper investigates time-varying decentralized feedback control architectures based on hybrid zero dynamics (HZD) that stabilize dynamic legged locomotion with high degrees of freedom. By conforming to the natural symmetries present in the robot's full-order model, three decentralization schemes are proposed for control synthesis, namely left-right, front-hind and diagonal. Our approach considers the strong nonlinear interactions between the subsystems and relies only on the intrinsic communication of the body's translation and rotational data that is readily available. Further, a quadratic programming (QP) based feedback linearization is employed to compute the control inputs for each subsystem. The effectiveness of the HZD-based decentralization scheme is demonstrated numerically for the stabilization of forward and inplace walking gaits on an 18 DOF robot.

This paper presents a new coordination algorithm for decentralised multi-robot information gathering. We consider planning for an online variant of the multi-agent orienteering problem with neighbourhoods. This formulation closely aligns with a number of important tasks in robotics, including inspection, surveillance, and reconnaissance. We propose a decentralised variant of the self-organising map (SOM) learning procedure, named Dec-SOM, which efficiently plans sequences of waypoints for a team of robots. Decentralisation is achieved by performing a distributed allocation scheme jointly with a series of SOM adaptations. We also offer an efficient heuristic to select when to perform negotiations, which reduces communication resource usage. Simulation results in two settings, including an infrastructure inspection scenario with a real-world dataset of oil rigs, demonstrate that Dec-SOM outperforms baseline methods and other SOM variants, is competitive with centralised SOM, and is a viable solution for decentralised information gathering.

Enabling multiple robots to collaboratively per-form coverage path planning on complex surfaces embedded in {{\mathbb{R}}^3}{{\mathbb{R}}^3} in the presence of moving obstacles is a challenging problem that has not received much attention from researchers. As robots start to be practically deployed, it is becoming important to address this problem. A novel decentralized multi-robot coverage path planning approach is proposed that is adaptive to unexpected stationary and moving obstacles while aiming to achieve complete coverage with minimal cost. The approach is inspired by the predator-prey relation. For a robot (a prey), a virtual stationary predator enforces spatial ordering on the prey, and dynamic predators (other robots) cause the prey to be repelled resulting in better task allocation and collision-avoidance. The approach makes the best use of both worlds: offline global planning for tuning of model parameters based on a prior map of the surface, and real-time local planning for adaptive and swift decision making amid moving obstacles and other robots while preserving global behavior. Comparisons with other approaches and extensive testing and validation using different number of robots, different surfaces and obstacles, and various scenarios are conducted.

As the number of robots in our daily surroundings like home, office, restaurants, factory floors, etc. are increasing rapidly, the development of natural human-robot interaction mechanism becomes more vital as it dictates the usability and acceptability of the robots. One of the valued features of such a cohabitant robot is that it performs tasks that are instructed in natural language. However, it is not trivial to execute the human intended tasks as natural language expressions can have large linguistic variations. Existing works assume either single task instruction is given to the robot at a time or there are multiple independent tasks in an instruction. However, complex task instructions composed of multiple inter-dependent tasks are not handled efficiently in the literature. There can be ordering dependency among the tasks, i.e., the tasks have to be executed in a certain order or there can be execution dependency, i.e., input parameter or execution of a task depends on the outcome of another task. Understanding such dependencies in a complex instruction is not trivial if an unconstrained natural language is allowed. In this work, we propose a method to find the intended order of execution of multiple inter-dependent tasks given in natural language instruction. Based on our experiment, we show that our system is very accurate in generating a viable execution plan from a complex instruction.

Nowadays, the prevalence of sensor networks has enabled tracking of the states of dynamic objects for a wide spectrum of applications from autonomous driving to environmental monitoring and urban planning. However, tracking realworld objects often faces two key challenges: First, due to the limitation of individual sensors, state estimation needs to be solved in a collaborative and distributed manner. Second, the objects' movement behavior model is unknown, and needs to be learned using sensor observations. In this work, for the first time, we formally formulate the problem of simultaneous state estimation and behavior learning in a sensor network. We then propose a simple yet effective solution to this new problem by extending the Gaussian process-based Bayes filters (GPBayesFilters) to an online, distributed setting. The effectiveness of the proposed method is evaluated on tracking objects with unknown movement behaviors using both synthetic data and data collected from a multi-robot platform.

This paper proposes a data-driven method for powered prosthesis control that achieves stable walking without the need for additional sensors on the human. The key idea is to extract the nominal gait and the human interaction information from motion capture data, and reconstruct the walking behavior with a dynamic model of the human-prosthesis system. The walking behavior of a human wearing a powered prosthesis is obtained through motion capture, which yields the limb and joint trajectories. Then a nominal trajectory is obtained by solving a gait optimization problem designed to reconstruct the walking behavior observed by motion capture. Moreover, the interaction force profiles between the human and the prosthesis are recovered by simulating the model following the recorded gaits, which are then used to construct a force tube that covers all the interaction force profiles. Finally, a robust Control Lyapunov Function (CLF) Quadratic Programming (QP) controller is designed to guarantee the convergence to the nominal trajectory under all possible interaction forces within the tube. Simulation results show this controller's improved tracking performance with a perturbed force profile compared to other control methods with less model information.

In this paper, our motivating application lies in precision agriculture where accurate modeling of forage is essential for informing rotational grazing strategies. Unfortunately, a major difficulty arises in modeling forage processes as they evolve on large scales according to complex ecological influences. As robots can collect data over large scales in a forage environment, they act as a promising resource for the forage modeling problem when combined with a data-driven Gaussian processes (GPs) technique. However, GPs are nonparametric in nature and may be blind to certain nuances of a process that a parameterized expert model may predict well. Indeed, for the forage modeling problem specifically, there exist several highly parameterized models from agricultural experts that exhibit powerful predictive capabilities. Expert models, however, often come with two shortcomings: (1) parameters may be difficult to determine in general; and (2) the model may not make complete spatiotemporal predictions. For example, a stochastic differential equation (SDE) that models the dynamics of the average output of an environment may be available from experts (a typical case). In such cases, we propose to take advantage of both data-driven (GPs) and expert (SDE) models, by fusing data collected by robots, which often yields spatial insight, with models from experienced professionals that often yield temporal insights. Specifically, we propose to leverage Bayesian estimation to combine these two methods, resulting in a posterior prediction that is a hybrid of data-driven and expert models. Finally, we provide simulations to demonstrate the effectiveness of the proposed method.

As electric vehicle (EV) technologies become mature, EV has been rapidly adopted in modern transportation systems, and is expected to provide future autonomous mobility-on-demand (AMoD) service with economic and societal benefits. However, EVs require frequent recharges due to their limited and unpredictable cruising ranges, and they have to be managed efficiently given the dynamic charging process. It is urgent and challenging to investigate a computationally efficient algorithm that provide EV AMoD system performance guarantees under model uncertainties, instead of using heuristic demand or charging models. To accomplish this goal, this work designs a data-driven distributionally robust optimization approach for vehicle supply-demand ratio and charging station utilization balancing, while minimizing the worst-case expected cost considering both passenger mobility demand uncertainties and EV supply uncertainties. We then derive an equivalent computationally tractable form for solving the distributionally robust problem in a computationally efficient way under ellipsoid uncertainty sets constructed from data. Based on E-taxi system data of Shenzhen city, we show that the average total balancing cost is reduced by 14.49%, the average unfairness of supply-demand ratio and utilization is reduced by 15.78% and 34.51% respectively with the distributionally robust vehicle balancing method, compared with solutions which do not consider model uncertainties.