Legged locomotion over various terrains is challenging and requires precise perception of the robot and its surroundings from both proprioception and vision. However, learning directly from high-dimensional visual input is often data-inefficient and intricate. To address this issue, traditional methods attempt to learn a teacher policy with access to privileged information first and then learn a student policy to imitate the teacher's behavior with visual input. Despite some progress, this imitation framework prevents the student policy from achieving optimal performance due to the information gap between inputs. Furthermore, the learning process is unnatural since animals intuitively learn to traverse different terrains based on their understanding of the world without privileged knowledge. Inspired by this natural ability, we propose a simple yet effective method, World Model-based Perception (WMP), which builds a world model of the environment and learns a policy based on the world model. We illustrate that though completely trained in simulation, the world model can make accurate predictions of real-world trajectories, thus providing informative signals for the policy controller. Extensive simulated and real-world experiments demonstrate that WMP outperforms state-of-the-art baselines in traversability and robustness. Videos and Code are available at: https://wmp-loco.github.io/.

In recent years, with the development of ship intelligence and uncrewed technology, autonomous path planning of uncrewed surface vehicle (USV) has become very important. However, most of the existing studies ignore the limitation of the endurance and speed of USV to voyage, and only consider the course planning at the fixed speed. To this end, we propose a path planning method based on deep reinforcement learning (DRL): first, we establish a USV motion model. Then, we propose a new USV energy consumption (EC) model. Existing research typically builds USV EC model by considering the combined effects of currents, winds, and waves, which needs precise ocean data and involves a complex modeling process. Moreover, it fail to account for the influence of USV’s own speed on EC. In contrast, based on the relationship between speed, marine environment, and propulsion load, we propose a new USV EC model. This model simplifies the modeling process and links USV EC with sailing speed and ocean environmental conditions. Next, the original soft-actor-critic (SAC) algorithm use the multilayer perceptron (MLP) as the action network. However, convolutional neural networks (CNNs) excel in capturing spatial and local features. Compared to MLP, CNN have stronger information acquisition capabilities. Therefore, we propose a model (FRCF) that combines fully connected layers with CNN to replace MLP as the action network of the SAC agent, aiming to enhance the agent’s convergence speed and performance. Finally, utilizing SAC with FRCF as the action network (SAC-FRCF), along with the motion models and the new USV EC model, we achieved multiobjective cooperative intelligent path planning for energy, speed, and heading. Unlike traditional path planning methods that control the USV heading based on discrete heading angular, we adjust USV speed and heading based on continuous acceleration and angular velocity. Meanwhile, we impose constraints on the USV’s speed, heading angle, acceleration, and angular velocity to ensure that its motion complies with kinematic constraints. Experimental results show that SAC-FRCF reduces the exploration time by 46.40% and has stronger convergence performance as well as better path planning effect compared with SAC algorithm using MLP as action network (SAC-MLP). Compared with SAC-MLP without considering the optimal EC, it reduces the EC by 28.38%. The experimental results verify the effectiveness and superiority of the proposed method. Furthermore, the method we propose exhibits significant advantages over traditional path planning methods. In addition, our proposed method demonstrates significant superiority compared to existing path planning methods, and its robustness and adaptability also have been verified in a more complex environment.

Wheeled robots have gained significant attention due to their wide range of applications in manufacturing, logistics, and service industries. However, due to the difficulty of building a highly accurate dynamics model for wheeled robots, developing and testing control algorithms for them remains challenging and time-consuming, requiring extensive physical experimentation. To address this problem, we propose D4W, i.e., Dependable Data-Driven Dynamics for Wheeled Robots, a simulation framework incorporating data-driven methods to accelerate the development and evaluation of algorithms for wheeled robots. The key contribution of D4W is a solution that utilizes real-world sensor data to learn accurate models of robot dynamics. The learned dynamics can capture complex robot behaviors and interactions with the environment throughout simulations, surpassing the limitations of analytical methods, which only work in simplified scenarios. Experimental results show that D4W achieves the best simulation accuracy compared to traditional approaches, allowing for rapid iteration of wheel robot algorithms with less or no need for fine-tuning in reality. We further verify the usability and practicality of the proposed framework through integration with existing simulators and controllers.

By transferring both features and gradients between different layers, shortcut connections explored by ResNets allow us to effectively train very deep neural networks up to hundreds of layers. However, the additional computation costs induced by those shortcuts are often overlooked. For example, during online inference, the shortcuts in ResNet-50 account for about 40 percent of the entire memory usage on feature maps, because the features in the preceding layers cannot be released until the subsequent calculation is completed. In this work, for the first time, we consider training the CNN models with shortcuts and deploying them without. In particular, we propose a novel joint-training framework to train plain CNN by leveraging the gradients of the ResNet counterpart. During forward step, the feature maps of the early stages of plain CNN are passed through later stages of both itself and the ResNet counterpart to calculate the loss. During backpropagation, gradients calculated from a mixture of these two parts are used to update the plainCNN network to solve the gradient vanishing problem. Extensive experiments on ImageNet/CIFAR10/CIFAR100 demonstrate that the plainCNN network without shortcuts generated by our approach can achieve the same level of accuracy as that of the ResNet baseline while achieving about $1.4\times $ speed-up and $1.25\times$ memory reduction. We also verified the feature transferability of our ImageNet pretrained plain-CNN network by fine-tuning it on MIT 67 and Caltech 101. Our results show that the performance of the plain-CNN is slightly higher than that of its baseline ResNet-50 on these two datasets. The codes are in: \href{https://github.com/leoozy/JointRD_Neurips2020}{https://github.com/leoozy/JointRD\_Neurips2020}

SUMMARY Reducing the time that a user has to occupy resources for completing cloud tasks can improve cloud efficiency and lower user cost. Such a time, called cloud time, consists of cloud deployment time and application running time. In this work, we design Jump‐start cloud, under which an efficient cloud deployment scheme is proposed for minimizing cloud time. It is especially beneficial to minimize cloud time for clouds serving multimedia applications, as they occupy much more significant cloud resources, such as disk I/O and networking bandwidth and computational costs in processing or transcoding multimedia contents. In particular, we have implemented virtual machine cloning based on disk image sharing for fast virtual machine and application deployment. For applications with heavy disk visits, the post‐deployment QoS may suffer from image sharing, and consequently, application running time will increase. To solve this problem, different image distribution schemes have been designed. We test Jump‐start cloud through a Hadoop‐based benchmark and MapReduce applications. Experiment studies show that our design saves application installation time and, meanwhile, keeps application running time reasonably low, thus makes cloud time shorter. Copyright © 2011 John Wiley & Sons, Ltd.

Opportunistic routing significantly increases unicast throughput in wireless mesh networks by effectively utilizing the wireless broadcast medium. With network coding, opportunistic routing can be implemented in a simple and practical way without resorting to a complicated scheduling protocol. Traditionally, due to the constraints of computational complexity, a protocol utilizing network coding needs to partition the data into multiple segments and encode only packets in the same segment. However, it is extremely challenging to decide the optimal time to move to the transmissions of the next segment, and existing designs all resort to different heuristic ideas that might harm network throughput. To address this problem, we proposeSlideOR, a new protocol to encode source packets in overlapping sliding windows such that coded packets from one window position may be useful towards decoding the source packets inside another window position. Through extensive simulations, we show thatSlideORoutperforms the existing solutions and is amenable to much simpler implementation than solutions with complicated scheduling among multiple segments.

The highly stochastic nature of wireless environments makes it desirable to monitor link loss rates in wireless sensor networks. In this paper, we study the loss inference problem in sensor networks with network coding. Unlike traditional transmission protocols, network coding offers reliable communication without using control messages for individual packets. We show, however, that network coding changes the fundamental connection between path and link loss probabilities such that new inference algorithms need to be developed. As end- to-end data are not sufficient to compute link loss rates precisely, we propose inference algorithms based on Bayesian principles to discover the set of highly lossy links in sensor networks. We show that our algorithms achieve high detection and low false-positive rates through extensive simulations.

Node churn and failures exist as fundamental characteristics in both peer-to-peer (P2P) and sensor networks. Peers in P2P networks are highly dynamic, whereas sensors are not dependable. As such, maintaining the persistence of periodically measured data in a scalable fashion has become a critical challenge in such systems, without the use of centralized servers. To better cope with node dynamics and failures, we propose priority random linear codes (RLCs), as well as their affiliated predistribution protocols, to maintain measurement data in different priorities, such that critical data have a higher opportunity to survive node failures than data of less importance. A salient feature of priority RLCs is the ability to partially recover more important subsets of the original data with higher priorities, when it is not feasible to recover all of them due to node dynamics. We present extensive analytical and experimental results to show the effectiveness of priority RLCs.

Wireless sensor networks consist of unreliable and energy-constrained sensors connecting to each other wirelessly. As measured data may be lost due to sensor failures, maintaining the persistence of periodically measured data in a scalable fashion has become a critical challenge in sensor networks, without the use of centralized servers. To cope with node failures, while providing convenient access to measured data, we propose geometric random linear codes, to encode data in a hierarchical fashion in geographic regions with different sizes, such that data are easy to access, if the original sensors producing the data are alive. Otherwise, data are persistently available elsewhere in the network. Although our coding scheme is simple, we have shown that it enjoys the same low encoding cost as sparse random linear codes, while dramatically decreasing the decoding cost. We present extensive analytical and experimental results to show the effectiveness of geometric random linear codes.

Opportunistic routing significantly increases unicast throughput in wireless mesh networks by effectively utilizing the wireless broadcast medium. With network coding, opportunistic routing can be implemented in a simple and practical way without resorting to a complicated scheduling protocol. Due to constraints of computational complexity, a protocol utilizing network coding needs to perform segmented network coding, which partitions the data into multiple segments and encode only packets in the same segment. However, existing designs transmit only one segment at any given time while waiting for its acknowledgment, which degrades performance as the size of the network scales up. In this paper, we propose CodeOR, a new protocol that uses network coding in opportunistic routing to improve throughput. By transmitting a window of multiple segments concurrently, it improves the performance of existing work by a factor of two on average (and a factor of four in some cases). CodeOR is especially appropriate for real-time multimedia applications through the use of a small segment size to decrease decoding delay, and is able to further increase network throughput with a smaller packet size and a larger window size.

Most routing protocols in disruption tolerant networks (DTN) use redundant transmissions to explore the diversities in routing paths in order to reduce data transmission delay. However, mobile nodes in DTN usually have limited energy and may prefer fewer transmissions for longer lifetime. Hence, it is vital to carefully balance the tradeoff between data transmission delay and the amount of transmissions among mobile nodes. In this paper, we consider the problem to route a batch of data packets in DTN. By making an analogy between the routing protocol and low-density erasure codes, we investigate the information-theoretical optimal number of data transmissions in delivering data. With such insights, we propose E-NCP, an efficient protocol in DTNs based on network coding, that reduces data transmissions significantly, while increasing data transmission delay only slightly as compared to the protocol with the best performance. With extensive theoretical analysis and simulations, we show that network coding facilitates a better tradeoff between resource usage and protocol performance, and that our protocol offers unique advantages over replication-based protocols.

Epidemic routing has been proposed to reduce the data transmission delay in disruption tolerant wireless networks, in which data can be replicated along multiple opportunistic paths as different nodes move within each other's communication range. With the advent of network coding, it is intuitive that data can not only be replicated, but also coded, when the transmission opportunity arises. However, will opportunistic communication with network coding perform any better than simple replications? In this paper, we present a stochastic analytical framework to study the performance of epidemic routing using network coding in opportunistic networks, as compared to the use of replication. We analytically show that network coding is superior when bandwidth and node buffers are limited, reflecting more realistic scenarios. Our analytical study is able to provide further insights towards future designs of efficient data communication protocols using network coding. As an example, we propose a priority based coding protocol, with which the destination can decode a high priority subset of the data much earlier than it can decode any data without the use of priorities. The correctness of our analytical results has also been confirmed by our extensive simulations.

Both peer-to-peer and sensor networks have the fundamental characteristics of node churn and failures. Peers in P2P networks are highly dynamic, whereas sensors are not dependable. As such, maintaining the persistence of periodically measured data in a scalable fashion has become a critical challenge in such systems, without the use of centralized servers. To better cope with node dynamics and failures, we propose priority random linear codes, as well as their affiliated pre-distribution protocols, to maintain measurement data in different priorities, such that critical data have a higher opportunity to survive node failures than data of less importance. A salient feature of priority random linear codes is the ability to partially recover more important subsets of the original data with higher priorities, when it is not feasible to recover all of them due to node dynamics. We present extensive analytical and experimental results to show the effectiveness of priority random linear codes.

It may not be feasible for sensor networks monitoring nature and inaccessible geographical regions to include powered sinks with Internet connections. We consider the scenario where sinks are not present in large-scale sensor networks, and unreliable sensors have to collectively resort to storing sensed data over time on themselves. At a time of convenience, such cached data from a small subset of live sensors may be collected by a centralized (possibly mobile) collector. In this paper, we propose a decentralized algorithm using fountain codes to guarantee the persistence and reliability of cached data on unreliable sensors. With fountain codes, the collector is able to recover all data as long as a sufficient number of sensors are alive. We use random walks to disseminate data from a sensor to a random subset of sensors in the network. Our algorithms take advantage of the low decoding complexity of fountain codes, as well as the scalability of the dissemination process via random walks. We have proposed two algorithms based on random walks. Our theoretical analysis and simulation-based studies have shown that, the first algorithm maintains the same level of fault tolerance as the original centralized fountain code, while introducing lower overhead than naive random-walk based implementation in the dissemination process. Our second algorithm has lower level of fault tolerance than the original centralized fountain code, but consumes much lower dissemination cost.

Wireless networks have been rapidly developing in recent years. General Packet Radio Service (GPRS) and Code Division Multiple Access (CDMA 1X) for wide areas, and 802.11 and Bluetooth for local areas have already emerged. Broadband wireless networks urgently call for rich contents for consumers. Among various possible applications, video communication is one of the most promising for mobile devices on wireless networks. This paper describes the generation, coding, and transmission of an effective video form, scalable portrait video for mobile video communication. As an expansion to bilevel video, portrait video is composed of more gray levels, and therefore possesses higher visual quality while it maintains a low bit rate and low computational costs. Portrait video is a scalable video in that each video with a higher level always contains all the information of the video with a lower level. The bandwidths of 2-4-level portrait videos fit into the bandwidth range of 20-40 kbps that GPRS and CDMA 1X can stably provide; therefore, portrait video is very promising for video broadcast and communication on 2.5-G wireless networks. With portrait video technology, we are the first to enable two-way video communication on pocket PCs and handheld PCs.