ICRA 2020 论文列表

Dynamic obstacle avoidance on quadrotors requires low latency. A class of sensors that are particularly suitable for such scenarios are event cameras. In this paper, we present a deep learning based solution for dodging multiple dynamic obstacles on a quadrotor with a single event camera and on-board computation. Our approach uses a series of shallow neural networks for estimating both the ego-motion and the motion of independently moving objects. The networks are trained in simulation and directly transfer to the real world without any fine-tuning or retraining. We successfully evaluate and demonstrate the proposed approach in many real-world experiments with obstacles of different shapes and sizes, achieving an overall success rate of 70% including objects of unknown shape and a low light testing scenario. To our knowledge, this is the first deep learning - based solution to the problem of dynamic obstacle avoidance using event cameras on a quadrotor. Finally, we also extend our work to the pursuit task by merely reversing the control policy, proving that our navigation stack can cater to different scenarios.

Inflated continuum robots are promising for a variety of navigation tasks, but controlling their motion with a small number of actuators is challenging. These inflated beam robots tend to buckle under compressive loads, producing extremely tight local curvature at difficult-to-control buckle point locations. In this paper, we present an inflated beam robot that uses distributed stiffness changing sections enabled by positive pressure layer jamming to control or prevent buckling. Passive valves are actuated by an electromagnet carried by an electromechanical device that travels inside the main inflated beam robot body. The valves themselves require no external connections or wiring, allowing the distributed stiffness control to be scaled to long beam lengths. Multiple layer jamming elements are stiffened simultaneously to achieve global stiffening, allowing the robot to support greater cantilevered loads and longer unsupported lengths. Local stiffening, achieved by leaving certain layer jamming elements unstiffened, allows the robot to produce "virtual joints" that dynamically change the robot kinematics. Implementing these stiffening strategies is compatible with growth through tip eversion and tendonsteering, and enables a number of new capabilities for inflated beam robots and tip-everting robots.

In order to fulfill conflicting requirements in the development of industrial robots, such as increased accuracy of a weightreduced manipulator with lower mechanical stiffness, the robot's dynamical behavior must be evaluated early in the development process. This leads to the need of accurate multibody models of the manipulator under development.This paper deals with multibody models that include flexible bodies, which are exported from the corresponding Finite Element model of the structural parts. It is shown that such a flexible link manipulator model, which is purely based on development and datasheet data, is suitable for an accurate description of an industrial robot's dynamic behavior. No stiffness parameters need to be identified by experimental methods, making this approach especially relevant during the development of new manipulators. This paper presents results of experiments in time and frequency domain for analyzing the modeling approach and for validating the model performance against real robot behavior.

The static world assumption is standard in most simultaneous localisation and mapping (SLAM) algorithms. Increased deployment of autonomous systems to unstructured dynamic environments is driving a need to identify moving objects and estimate their velocity in real-time. Most existing SLAM based approaches rely on a database of 3D models of objects or impose significant motion constraints. In this paper, we propose a new feature-based, model-free, object-aware dynamic SLAM algorithm that exploits semantic segmentation to allow estimation of motion of rigid objects in a scene without the need to estimate the object poses or have any prior knowledge of their 3D models. The algorithm generates a map of dynamic and static structure and has the ability to extract velocities of rigid moving objects in the scene. Its performance is demonstrated on simulated, synthetic and real-world datasets.

In this work, we propose an augmentation to the Dynamic Movement Primitives (DMP) framework which allows the system to generalize to moving goals without the use of any known or approximation model for estimating the goal's motion. We aim to maintain the demonstrated velocity levels during the execution to the moving goal, generating motion profiles appropriate for human robot collaboration. The proposed method employs a modified version of a DMP, learned by a demonstration to a static goal, with adaptive temporal scaling in order to achieve reaching of the moving goal with the learned kinematic pattern. Only the current position and velocity of the goal are required. The goal's reaching error and its derivative is proved to converge to zero via contraction analysis. The theoretical results are verified by simulations and experiments on a KUKA LWR4+ robot.

Autonomous landing on a moving platform presents unique challenges for multirotor vehicles, including the need to accurately localize the platform, fast trajectory planning, and precise/robust control. Previous works studied this problem but most lack explicit consideration of the wind disturbance, which typically leads to slow descents onto the platform. This work presents a fully autonomous vision-based system that addresses these limitations by tightly coupling the localization, planning, and control, thereby enabling fast and accurate landing on a moving platform. The platform's position, orientation, and velocity are estimated by an extended Kalman filter using simulated GPS measurements when the quadrotor-platform distance is large, and by a visual fiducial system when the platform is nearby. The landing trajectory is computed online using receding horizon control and is followed by a boundary layer sliding controller that provides tracking performance guarantees in the presence of unknown, but bounded, disturbances. To improve the performance, the characteristics of the turbulent conditions are accounted for in the controller. The landing trajectory is fast, direct, and does not require hovering over the platform, as is typical of most stateof-the-art approaches. Simulations and hardware experiments are presented to validate the robustness of the approach.

The common pipeline in autonomous driving systems is highly modular and includes a perception component which extracts lists of surrounding objects and passes these lists to a high-level decision component. In this case, leveraging the benefits of deep reinforcement learning for high-level decision making requires special architectures to deal with multiple variable-length sequences of different object types, such as vehicles, lanes or traffic signs. At the same time, the architecture has to be able to cover interactions between traffic participants in order to find the optimal action to be taken. In this work, we propose the novel Deep Scenes architecture, that can learn complex interaction-aware scene representations based on extensions of either 1) Deep Sets or 2) Graph Convolutional Networks. We present the Graph-Q and DeepScene-Q off-policy reinforcement learning algorithms, both outperforming state-ofthe-art methods in evaluations with the publicly available traffic simulator SUMO.

This paper employs velocity decomposition of underactuated mechanical systems to determine the degree of dynamic coupling in the gaits of a two-link biped model. The degree of coupling between controlled and uncontrolled directions quantifies the control authority the system has over its unactuated degree of freedom. This paper shows that the amount of coupling is directly correlated to gait robustness, as seen through the size of the gait's region of attraction. The analytical measure of coupling is applied in the context of trajectory optimization to generate two-link gaits that maximize or minimize coupling. Simulation studies show that gaits maximizing coupling exhibit significantly superior robustness, as measured by 1) stochastic performance on uneven terrain, 2) ability to maintain desired walking speed under non-vanishing disturbances, 3) size of the region of attraction, and 4) robustness to model uncertainties.

In this paper we present a Deep Reinforcement Learning approach to solve dynamic cloth manipulation tasks. Differing from the case of rigid objects, we stress that the followed trajectory (including speed and acceleration) has a decisive influence on the final state of cloth, which can greatly vary even if the positions reached by the grasped points are the same. We explore how goal positions for non-grasped points can be attained through learning adequate trajectories for the grasped points. Our approach uses few demonstrations to improve control policy learning, and a sparse reward approach to avoid engineering complex reward functions. Since perception of textiles is challenging, we also study different state representations to assess the minimum observation space required for learning to succeed. Finally, we compare different combinations of control policy encodings, demonstrations, and sparse reward learning techniques, and show that our proposed approach can learn dynamic cloth manipulation in an efficient way, i.e., using a reduced observation space, a few demonstrations, and a sparse reward.

Anchor boxes act as potential object localization candidates allow single-stage detectors to achieve real-time performance, at the cost of localization accuracy when compared to state-of-the-art two-stage detectors. Therefore, correct selection of the scale and aspect ratio associated with an anchor box is crucial for detector performance. In this work, we propose a novel architecture called DANet for improving the localization performance of single-stage object detectors, while maintaining real-time inference. The proposed network achieves this by predicting (1) the combination of aspect ratio and scale per feature map based on object density and (2) localization confidence per anchor box. We evaluate the proposed network using the benchmark dataset. On the MS COCO dataset, DANet achieves 30.9% AP at 51.8 fps using ResNet-18 and 45.3% AP at 7.4 fps using ResNeXt-101. The code and models will be available at https://github.com/PS06/AnchorNet.

Real-world robots have complex strict constraints. Therefore, safe reinforcement learning algorithms that can simultaneously minimize the total cost and the risk of constraint violation are crucial. However, almost no algorithms exist that can scale to high-dimensional systems to the best of our knowledge. In this paper, we propose Dynamic Actor-Advisor Programming (DAAP), as an algorithm for sample-efficient and scalable safe reinforcement learning. DAAP employs two control policies, actor and advisor. They are updated to minimize total cost and risk of constraint violation intertwiningly and smoothly towards each other's direction by using the other as the baseline policy in the Kullback-Leibler divergence of Dynamic Policy Programming framework. We demonstrate the scalability and sample efficiency of DAAP through its application on simulated robot arm control tasks with performance comparisons to baselines.

Besides being part of the Internet of Things (IoT), drones can play a relevant role in it as enablers. The 3D mobility of UAVs can be exploited to improve node localization in IoT networks for, e.g., search and rescue or goods localization and tracking. One of the widespread IoT communication technologies is Long Range Wide Area Network (LoRaWAN), which allows achieving long communication distances with low power. In this work, we present a drone-aided localization system for LoRa networks in which a UAV is used to improve the estimation of a node's location initially provided by the network. We characterize the relevant parameters of the communication system and use them to develop and test a search algorithm in a realistic simulated scenario. We then move to the full implementation of a real system in which a drone is seamlessly integrated into Swisscom's LoRa network. The drone coordinates with the network with a two-way exchange of information which results in an accurate and fully autonomous localization system. The results obtained in our field tests show a ten-fold improvement in localization precision with respect to the estimation provided by the fixed network. Up to our knowledge, this is the first time a UAV is successfully integrated in a LoRa network to improve its localization accuracy.

Traditional planning and control methods could fail to find a feasible trajectory for an autonomous vehicle to execute amongst dense traffic on roads. This is because the obstacle-free volume in spacetime is very small in these scenarios for the vehicle to drive through. However, that does not mean the task is infeasible since human drivers are known to be able to drive amongst dense traffic by leveraging the cooperativeness of other drivers to open a gap. The traditional methods fail to take into account the fact that the actions taken by an agent affect the behaviour of other vehicles on the road. In this work, we rely on the ability of deep reinforcement learning to implicitly model such interactions and learn a continuous control policy over the action space of an autonomous vehicle. The application we consider requires our agent to negotiate and open a gap in the road in order to successfully merge or change lanes. Our policy learns to repeatedly probe into the target road lane while trying to find a safe spot to move in to. We compare against two model-predictive control-based algorithms and show that our policy outperforms them in simulation. As part of this work, we introduce a benchmark for driving in dense traffic for use by the community.

General-purpose planning algorithms for automated driving combine mission, behavior, and local motion planning. Such planning algorithms map features of the environment and driving kinematics into complex reward functions. To achieve this, planning experts often rely on linear reward functions. The specification and tuning of these reward functions is a tedious process and requires significant experience. Moreover, a manually designed linear reward function does not generalize across different driving situations. In this work, we propose a deep learning approach based on inverse reinforcement learning that generates situation-dependent reward functions. Our neural network provides a mapping between features and actions of sampled driving policies of a model-predictive control-based planner and predicts reward functions for upcoming planning cycles. In our evaluation, we compare the driving style of reward functions predicted by our deep network against clustered and linear reward functions. Our proposed deep learning approach outperforms clustered linear reward functions and is at par with linear reward functions with a-priori knowledge about the situation.

In a distributed multi-master-multi-slave teleoperation system, the human users may compete against each other for the control of the team of slave robots. To win the competition, one operator would send the largest command to the slave group. For the sake of team cohesion, the slave group should follow the command of the winning operator and ignore the commands of the other users. To enable (i) the slave team to identify the winning operator, and (ii) each slave to determine whether to admit or discard the command it receives from its operator, this paper proposes a dynamic decision-making protocol that distinguishes the decision variable of the slave commanded by the winner from the decision variables of all other slave robots. The protocol only requires the slaves to exchange and evaluate their decision variables locally. Lyapunov stability analysis proves the theoretical convergence of the proposed decision-making algorithm. An experimental distributed winner-take-all teleoperation in a 3-masters-11-slaves teleoperation testbed validates its practical efficacy.

The RVM (Robot-based Vibration Suppression Modules) is proposed for the manipulation and transport of a large flexible object. Since the RVM is easily attachable/detachable to the object, this RVM allows distributing over the manipulated object so that it is scalable to the object size. The composition of the system is partly motivated by the MAGMaS (Multiple Aerial-Ground Manipulator System) [1]- [3], however, since the quadrotor usage is mechanically too complicated and its design is not optimized for manipulation, thus we overcome these limitations using distributed RVMs and newly developed theory. For this, we first provide a constrained optimization problem of RVM design with the minimum number of rotors, so that the feasible thrust force is maximized while it minimizes undesirable wrench and its own weight. Then, we derive the full dynamics and elucidate a controllability condition with multiple distributed RVMs and show that even if multiple, their structures turn out similar to [2] composed with a single quadrotor. We also elucidate the optimal placement of the RVM via the usage of controllability gramian which is not even alluded in [2] and established for the first time here. Experiments are performed to demonstrate the effectiveness of the proposed theory.

This paper presents a distributed optimization framework and its local utility design for communication-aware information gathering by mobile robotic sensor networks. The main idea of the optimization is that each robot decides based on its local utility that considers the decisions of other neighbor robots higher in a given hierarchy. The local utility is designed as conditional mutual information that captures sensing and communication properties. Sampling procedures using a specific measurement set and particle methods are applied to compute the designed utility, which allows nonlinear, non-Gaussian properties of targets, sensing, and communication. Simulation results describe the presented distributed optimization shows more improved estimates and entropy reduction than another approach that does not consider communication properties. Simulation results also verify the presented distributed optimization using the described approach for information computation has better results than using other approaches that simplify the communication-aware information.

We present a scalable distributed target tracking algorithm based on the alternating direction method of multipliers that is well-suited for a fleet of autonomous cars communicating over a vehicle-to-vehicle network. Each sensing vehicle communicates with its neighbors to execute iterations of a Kalman filter-like update such that each agent's estimate approximates the centralized maximum a posteriori estimate without requiring the communication of measurements. We show that our method outperforms the Consensus Kalman Filter in recovering the centralized estimate given a fixed communication bandwidth. We also demonstrate the algorithm in a high fidelity urban driving simulator (CARLA), in which 50 autonomous cars connected on a time-varying communication network track the positions and velocities of 50 target vehicles using on-board cameras.

In this paper, we investigate the consensus problem of multiple unmanned aerial vehicles (UAVs) in the presence of environmental constraints under a general communication topology containing a directed spanning tree. First, based on a position transformation function, we propose a novel dynamic reference position and yaw angle for each UAV to cope with both the asymmetric topology and the constraints. Then, the backstepping-like design methodology is presented to derive a local tracking controller for each UAV such that its position and yaw angle can converge to the reference ones. The proposed protocol is distributed in the sense that, the input update of each UAV dynamically relies only on local state information from its neighborhood set and the constraints, and it does not require any additional centralized information. It is demonstrated that under the proposed protocol, all UAVs reach consensus without violation of the environmental constraints. Finally, simulation and experimental results are provided to demonstrate the performance of the protocol.