Environment feedback t = 0 t = 1 t = 2 t = 3 t = 4 t = H … Action Collision with wall Collision with Furniture No Collision Different Platforms time Training entirely in simulation Test in real world Fig. 1. We propose the Collision Avoidance via Deep Reinforcement Learning algorithm for indoor flight which is entirely trained in a simulated CAD environment. Left: CAD 2 RL uses single image inputs from a monocular camera, is exclusively trained in simulation, and does not see any real images at training time. Training is performed using a Monte Carlo policy evaluation method, which performs rollouts for multiple actions from each initial state and trains a deep network to predict long-horizon collision probabilities of each action. Right: CAD 2 RL generalizes to real indoor flight. Abstract-Deep reinforcement learning has emerged as a promising and powerful technique for automatically acquiring control policies that can process raw sensory inputs, such as images, and perform complex behaviors. However, extending deep RL to real-world robotic tasks has proven challenging, particularly in safety-critical domains such as autonomous flight, where a trial-and-error learning process is often impractical. In this paper, we explore the following question: can we train vision-based navigation policies entirely in simulation, and then transfer them into the real world to achieve real-world flight without a single real training image? We propose a learning method that we call CAD 2 RL, which can be used to perform collision-free indoor flight in the real world while being trained entirely on 3D CAD models. Our method uses single RGB images from a monocular camera, without needing to explicitly reconstruct the 3D geometry of the environment or perform explicit motion planning. Our learned collision avoidance policy is represented by a deep convolutional neural network that directly processes raw monocular images and outputs velocity commands. This policy is trained entirely on simulated images, with a Monte Carlo policy evaluation algorithm that directly optimizes the network's ability to produce collision-free flight. By highly randomizing the rendering settings for our simulated training set, we show that we can train a policy that generalizes to the real world, without requiring the simulator to be particularly realistic or high-fidelity. We evaluate our method by flying a real quadrotor through indoor environments, and further evaluate the design choices in our simulator through a series of ablation studies on depth prediction. For supplementary video see: https://youtu.be/nXBWmzFrj5s
translated by 谷歌翻译
How do you learn to navigate an Unmanned Aerial Vehicle (UAV) and avoid obstacles? One approach is to use a small dataset collected by human experts: however, high capacity learning algorithms tend to overfit when trained with little data. An alternative is to use simulation. But the gap between simulation and real world remains large especially for perception problems. The reason most research avoids using large-scale real data is the fear of crashes! In this paper, we propose to bite the bullet and collect a dataset of crashes itself! We build a drone whose sole purpose is to crash into objects: it samples naive trajectories and crashes into random objects. We crash our drone 11,500 times to create one of the biggest UAV crash dataset. This dataset captures the different ways in which a UAV can crash. We use all this negative flying data in conjunction with positive data sampled from the same trajectories to learn a simple yet powerful policy for UAV navigation. We show that this simple self-supervised model is quite effective in navigating the UAV even in extremely cluttered environments with dynamic obstacles including humans. For supplementary video see: https://youtu.be/u151hJaGKUo
translated by 谷歌翻译
We consider the task of driving a remote control car at high speeds through unstructured outdoor environments. We present an approach in which supervised learning is first used to estimate depths from single monoc-ular images. The learning algorithm can be trained either on real camera images labeled with ground-truth distances to the closest obstacles , or on a training set consisting of synthetic graphics images. The resulting algorithm is able to learn monocular vision cues that accurately estimate the relative depths of obstacles in a scene. Reinforcement learn-ing/policy search is then applied within a simulator that renders synthetic scenes. This learns a control policy that selects a steering direction as a function of the vision system's output. We present results evaluating the predictive ability of the algorithm both on held out test data, and in actual autonomous driving experiments.
translated by 谷歌翻译
Mobile robot vision-based navigation has been the source of countless research contributions, from the domains of both vision and control. Vision is becoming more and more common in applications such as localization, automatic map construction, autonomous navigation, path following, inspection, monitoring or risky situation detection. This survey presents those pieces of work, from the nineties until nowadays, which constitute a wide progress in visual navigation techniques for land, aerial and autonomous underwater vehicles. The paper deals with two major approaches: map-based navigation and mapless navigation. Map-based navigation has been in turn subdivided in metric map-based navigation and topological map-based navigation. Our outline to mapless navigation includes reactive techniques based on qualitative characteristics extraction, appearance-based localization, optical flow, features tracking, plane ground detection/tracking, etc... The recent concept of visual sonar has also been revised.
translated by 谷歌翻译
在人类活动示范训练的深度神经网络赋予了在路上进行自驾车的能力。然而,在诸如校园设施之类的行人富裕环境中导航仍然具有挑战性 - 需要经常对机器人进行干预并且从早期步骤控制机器人导致错误。因此,艰巨的负担放在学习框架设计和数据采集上。在本文中,我们提出了一种新的干预学习数据集聚(DAgger)算法,以克服在行人富裕环境中应用模仿学习导航所带来的局限性。我们的新学习算法实现了错误回溯功能,能够有效地学习fromexpert干预。结合我们的新学习算法与深度卷积神经网络和分层嵌套的策略选择机制,我们展示了我们的机器人能够将像素直接映射到控制命令并在现实世界中成功导航,而无需对行人行为或世界模型进行明确建模。
translated by 谷歌翻译
(a) Aerial view of test environment (b) Vision-based driving, view from onboard camera (c) Side view of vehicle Fig. 1. Conditional imitation learning allows an autonomous vehicle trained end-to-end to be directed by high-level commands. (a) We train and evaluate robotic vehicles in the physical world (top) and in simulated urban environments (bottom). (b) The vehicles drive based on video from a forward-facing onboard camera. At the time these images were taken, the vehicle was given the command "turn right at the next intersection". (c) The trained controller handles sensorimotor coordination (staying on the road, avoiding collisions) and follows the provided commands. Abstract-Deep networks trained on demonstrations of human driving have learned to follow roads and avoid obstacles. However, driving policies trained via imitation learning cannot be controlled at test time. A vehicle trained end-to-end to imitate an expert cannot be guided to take a specific turn at an up-coming intersection. This limits the utility of such systems. We propose to condition imitation learning on high-level command input. At test time, the learned driving policy functions as a chauffeur that handles sensorimotor coordination but continues to respond to navigational commands. We evaluate different architectures for conditional imitation learning in vision-based driving. We conduct experiments in realistic three-dimensional simulations of urban driving and on a 1/5 scale robotic truck that is trained to drive in a residential area. Both systems drive based on visual input yet remain responsive to high-level navigational commands.
translated by 谷歌翻译
In this paper, we present our latest achievements towards the goal of autonomous flights of an MAV in unknown environments, only having a monocular camera as exteroceptive sensor. As MAVs are highly agile, it is not sufficient to directly use the visual input for position control at the framerates that can be achieved with small onboard computers. Our contributions in this work are twofold. First, we present a solution to overcome the issue of having a low frequent onboard visual pose update versus the high agility of an MAV. This is solved by filtering visual information with inputs from inertial sensors. Second, as our system is based on monocular vision, we present a solution to estimate the metric visual scale aid of an air pressure sensor. All computation is running onboard and is tightly integrated on the MAV to avoid jitter and latencies. This framework enables stable flights indoors and outdoors even under windy conditions.
translated by 谷歌翻译
Search and Rescue (SAR) missions represent an important challenge in the robotics research field as they usually involve exceedingly variable-nature scenarios which require a high-level of autonomy and versatile decision-making capabilities. This challenge becomes even more relevant in the case of aerial robotic platforms owing to their limited payload and computational capabilities. In this paper, we present a fully-autonomous aerial robotic solution, for executing complex SAR missions in unstructured indoor environments. The proposed system is based on the combination of a complete hardware configuration and a flexible system architecture which allows the execution of high-level missions in a fully unsupervised manner (i.e. without human intervention). In order to obtain flexible and versatile behaviors from the proposed aerial robot, several learning-based capabilities have been integrated for target recognition and interaction. The target recognition capability includes a supervised learning classifier based on a computationally-efficient Convolutional Neural Network (CNN) model trained for target/background classification, while the capability to interact with the target for rescue operations introduces a novel Image-Based Visual Servoing (IBVS) algorithm which integrates a recent deep reinforcement learning method named Deep Deterministic Policy Gradients (DDPG). In order to train the aerial robot for performing IBVS tasks, a reinforcement learning framework has been developed, which integrates a deep reinforcement learning agent (e.g. DDPG) with a Gazebo-based simulator for aerial robotics. The proposed system has been validated in a wide range of simulation flights, using Gazebo and PX4 Software-In-The-Loop, and real flights in cluttered indoor environments, demonstrating the versatility of the proposed system in complex SAR missions.
translated by 谷歌翻译
Learning from demonstration for motion planning is an ongoing research topic. In this paper we present a model that is able to learn the complex mapping from raw 2D-laser range findings and a target position to the required steering commands for the robot. To our best knowledge, this work presents the first approach that learns a target-oriented end-to-end navigation model for a robotic platform. The supervised model training is based on expert demonstrations generated in simulation with an existing motion planner. We demonstrate that the learned navigation model is directly transferable to previously unseen virtual and, more interestingly, real-world environments. It can safely navigate the robot through obstacle-cluttered environments to reach the provided targets. We present an extensive qualitative and quantitative evaluation of the neural network-based motion planner, and compare it to a grid-based global approach, both in simulation and in real-world experiments.
translated by 谷歌翻译
We present a novel end-to-end learning framework to enable ground vehicles to autonomously navigate unknown environments by fusing raw pixels from a front facing camera and depth measurements from LiDAR. A new deep neural network architecture is introduced for mapping the depth and vision from LiDAR and camera, respectively, to the steering commands. The network effectively performs modality fusion and reliably predicts steering commands even in the presence of sensor failures. The proposed network in trained on our own dataset, which we will publicly release, of LiDAR depth measurements and camera images taken in an indoor corridor environment. Comprehensive experimental evaluation to demonstrate the robustness of our network architecture is performed to show that the proposed deep learning neural network is able to fully autonomously navigate in the corridor environment. Furthermore, we demonstrate that the fusion of the camera and LiDAR modalities provides further benefits beyond robustness to sensor failures. Specifically, the multimodal fused system shows a potential to navigate around obstacles placed in the corridor environment and to handle changes in environment geometry (e.g., having additional paths such as opening of doors that were closed during training) without being trained for these tasks.
translated by 谷歌翻译
大多数现有的自动驾驶方法属于两类:模块化管道,构建广泛的环境模型,以及模仿学习方法,直接映射图像控制输出。最近提出的第三种范式,即直接感知,旨在通过使用神经网络来学习适当的低维中间表示来实现两者的优点。然而,现有的直接感知方法仅限于简单的高速公路情况,缺乏导航交叉路口的能力,在交通信号灯处停下或遵守速度限制。在这项工作中,我们提出了一种直接感知方法,该方法将视频输入映射到适合自主导航的中间表示,给出高级方向输入。相比于最先进的强化和条件模仿学习方法,我们在具有挑战性的CARLA模拟基准测试中实现了目标导向导向高达68%的改进。此外,我们的方法是首先通过仅使用图像级标签来处理交通信号灯和速度标志,以及平稳的跟车,从而显着减少模拟中的交通事故。
translated by 谷歌翻译
The ability to transfer knowledge gained in previous tasks into new contexts is one of the most important mechanisms of human learning. Despite this, adapting autonomous behavior to be reused in partially similar settings is still an open problem in current robotics research. In this paper, we take a small step in this direction and propose a generic framework for learning transferable motion policies. Our goal is to solve a learning problem in a target domain by utilizing the training data in a different but related source domain. We present this in the context of an autonomous MAV flight using monocular reactive control, and demonstrate the efficacy of our proposed approach through extensive real-world flight experiments in outdoor cluttered environments.
translated by 谷歌翻译
本文介绍了我们的方法,使配备有单目相机的无人机四旋翼能够自主地避免与未构造和未知的室内环境中的障碍物碰撞。与地面车辆机器人中的障碍物相比,无人机导航带来了额外的挑战,因为无人机运动不再局限于明确的室内地面或街道环境。室内和室外环境中的水平结构,如装饰物品,家具,吊扇,标志牌,树枝等,也成为与地面车辆机器人不同的相关障碍。因此,为地面机器人开发的避障方法显然不适用于无人机导航。使用单眼图像用于无人机避障的当前控制方法严重依赖于环境信息。这些控制器不能完全保留和利用有关决策制定的周围环境的广泛可用信息。我们提出了一种基于深度强化学习的UAVobstacle避免(OA)和自主探索方法,它能够完全相同。我们方法中的关键思想是部分可观测性的概念以及无人机如何保留有关环境结构的相关信息,以便做出更好的未来导航决策。我们的OA技术使用具有时间关注度的递归神经网络,并且与先前的工作相比,在没有碰撞的导航期间覆盖的距离方面提供更好的结果。此外,我们的技术具有很高的推理力(机器人应用中的关键因素),并且能够高效节能,因为它最大限度地减少了无人机的振荡运动并减少了功耗。
translated by 谷歌翻译
RGB-D cameras provide both a color image and per-pixel depth estimates. The richness of their data and the recent development of low-cost sensors have combined to present an attractive opportunity for mobile robotics research. In this paper, we describe a system for visual odometry and mapping using an RGB-D camera, and its application to autonomous flight. By leveraging results from recent state-of-the-art algorithms and hardware, our system enables 3D flight in cluttered environments using only onboard sensor data. All computation and sensing required for local position control are performed onboard the vehicle, reducing the dependence on unreliable wireless links. We evaluate the effectiveness of our system for stabilizing and controlling a quadrotor micro air vehicle, demonstrate its use for constructing detailed 3D maps of an indoor environment, and discuss its limitations.
translated by 谷歌翻译
Mapping a river's geometry provides valuable information to help understand the topology and health of an environment and deduce other attributes such as which types of surface vessels could traverse the river. While many rivers can be mapped from satellite imagery, smaller rivers that pass through dense vegetation are occluded. We develop a micro air vehicle (MAV) that operates beneath the tree line, detects and maps the river, and plans paths around three-dimensional (3D) obstacles (such as overhanging tree branches) to navigate rivers purely with onboard sensing, with no GPS and no prior map. We present the two enabling algorithms for exploration and for 3D motion planning. We extract high-level goal-points using a novel exploration algorithm that uses multiple layers of information to maximize the length of the river that is explored during a mission. We also present an efficient modification to the SPARTAN (Sparse Tangential Network) algorithm called SPARTAN-lite, which exploits geodesic properties on smooth manifolds of a tangential surface around obstacles to plan rapidly through free space. Using limited onboard resources, the exploration and planning algorithms together compute trajectories through complex unstructured and unknown terrain, a capability rarely demonstrated by flying vehicles operating over rivers or over ground. We evaluate our approach against commonly employed algorithms and compare guidance decisions made by our system to those made by a human piloting a boat carrying our system over multiple kilometers. We also present fully autonomous flights on riverine environments generating 3D maps over several hundred-meter stretches of tight winding rivers. C 2015 Wiley Periodicals, Inc.
translated by 谷歌翻译
In this review, we provide an overview of emerging trends and challenges in the field of intelligent and autonomous, or self-driving, vehicles. Recent advances in the field of perception, planning, and decision-making for autonomous vehicles have led to great improvements in functional capabilities, with several prototypes already driving on our roads and streets. Yet challenges remain regarding guaranteed performance and safety under all driving circumstances. For instance, planning methods that provide safe and system-compliant performance in complex, cluttered environments while modeling the uncertain interaction with other traffic participants are required. Furthermore , new paradigms, such as interactive planning and end-to-end learning, open up questions regarding safety and reliability that need to be addressed. In this survey, we emphasize recent approaches for integrated perception and planning and for behavior-aware planning, many of which rely on machine learning. This raises the question of verification and safety, which we also touch upon. Finally, we discuss the state of the art and remaining challenges for managing fleets of autonomous vehicles. 8.1
translated by 谷歌翻译
自动驾驶的端到端方法具有高样本复杂性并且难以扩展到真实的城市驾驶。模拟可以通过提供廉价,安全和多样化的列车环境来实现驾驶系统。然而,在模拟中培训驾驶政策带来了将这些政策转移到现实世界的问题。我们提出了一种通过模块化和抽象将驱动策略从模拟转移到现实的方法。我们的方法受到经典驱动系统的启发,旨在实现模块化架构和端到端深度学习方法的优势。关键的想法是封装驾驶政策,使其不直接暴露于原始感知输入或低级车辆动态。我们在模拟城市环境和现实世界中评估所提出的方法。特别是,我们将在模拟中训练的驾驶政策转移到a1 / 5级机器人卡车,该卡车在各种条件下,在两个大陆上进行nofinetuning。补充视频可以通过以下网址查看://youtu.be/BrMDJqI6H5U
translated by 谷歌翻译
This paper describes the architecture and implementation of an autonomous passenger vehicle designed to navigate using locally perceived information in preference to potentially inaccurate or incomplete map data. The vehicle architecture was designed to handle the original DARPA Urban Challenge requirements of perceiving and navigating a road network with segments defined by sparse waypoints. The vehicle implementation includes many heterogeneous sensors with significant communications and computation bandwidth to capture and process high-resolution, high-rate sensor data. The output of the comprehensive environmental sensing subsystem is fed into a kinodynamic motion planning algorithm to generate all vehicle motion. The requirements of driving in lanes, three-point turns, parking, and maneuvering through obstacle fields are all generated with a unified planner. A key aspect of the planner is its use of closed-loop simulation in a rapidly exploring randomized trees algorithm, which can randomly explore the space while efficiently generating smooth trajectories in a dynamic and uncertain environment. The overall system was realized through the creation of a powerful new suite of software tools for message passing, logging, and visualization. These innovations provide a strong platform for future research in autonomous driving in global positioning system-denied and highly dynamic environments with poor a priori information. C 2008 Wiley Periodicals, Inc.
translated by 谷歌翻译
Automatic obstacle detection and avoidance is a key component for the success of micro-aerial vehicles (MAVs) in the future. As the payload of MAVs is highly constrained, cameras are attractive sensors because they are both lightweight and provide rich information about the environment. In this paper, we present an approach that allows a quadrotor with a single monocular camera to locally generate collision-free waypoints. We acquire a small set of images while the quadrotor is hovering from which we compute a dense depth map. Based on this depth map, we render a 2D scan and generate a suitable waypoint for navigation. In our experiments, we found that the pose variation during hovering is already sufficient to obtain suitable depth maps. The computation takes less than one second which renders our approach applicable for obstacle avoidance in real-time. We demonstrate the validity of our approach in challenging environments where we navigate a Parrot Ardrone quadrotor successfully through narrow passages including doors, boxes, and people.
translated by 谷歌翻译
This paper presents our solution for enabling a quadrotor helicopter to autonomously navigate, explore and locate objects of interest in unstructured and unknown indoor environments. We describe the design and operation of our quadrotor helicopter, before presenting the software architecture and individual algorithms necessary for executing the mission. Experimental results are presented demonstrating the quadrotor's ability to operate autonomously in indoor environments. Finally, we address some of the logistical and risk management issues pertaining to our entry in the competition.
translated by 谷歌翻译