Offline reinforcement learning is proposed to learn policy from static data. However, when the model is deployed to real world, models without safety constraints can cause unsafe problems in reality. Using negative rewards or restricting the agent's action space for unsafe policies in related work may lead to models being overly conservative or aggressive, thus failing to balance task performance and safety effectively. In this paper, we propose a safety offline reinforcement learning method based on knowledge constraints by adding prior expert knowledge from the static data set. First, our algorithm utilizes an expert model to evaluate safety checks on state-action pairs, aiming to ensure that the model learns from safe data. Second, we use adaptive adjustment factors to impose safety constraints on the current policy. When the current policy is evaluated unsafe, constraints on the policy are heightened. Conversely, when the policy is evaluated safety, the model optimizes task performance accordingly. Experiments demonstrate that our algorithm outperforms baseline algorithms in safety, exhibiting reduced variance and enhanced training stability.
In intelligent unmanned ground vehicle systems, decision-making algorithms often face challenges in adapting to dynamically changing environments, and their generalization capabilities may be limited. Most existing decision-making algorithms can only achieve robust results in the original scene, but when transferred to new scenes under same task, the algorithm performance drops sharply. Moreover, when deploying decision-making algorithms to unmanned ground vehicles, they often struggle to achieve performance comparable to computer simulations. To tackle this challenge, this paper proposes Scene Semantic Reconstruction for Unmanned Ground Vehicle Virtual-Real Integration(S2RU). S2RU decomposes the scene into abstract entities with object semantic information and then combines these entities using compositional neural radiance fields to enhance the capabilities of the UGV agent. This means the decision-making process is divided into two stages. In the first stage, concrete entities in the original perceptual information are mapped to abstract entities and transformed into scene semantic maps. In the second stage, decisions are made based on scene semantic maps. We have validated in both simulation and real-world environments, showcasing robust transferability between these environments and enabling cross-scene transfer for the same task and validate the usability, completeness and stability of S2RU.Results demonstrate that our methods improve success rate of a particular task across different scenes by at least 20% compared to other virtual-real integration methods.
Image stitching is one of the important tasks of computer vision, which is used in many fields such as autonomous navigation and autonomous driving. However, traditional stitching methods rely too much on the quality of feature detection and show poor performance for images with few features or low resolution. Although existing deep learning-based methods can make up for the shortcomings of traditional methods, they are only used on mobile robots with smallbaseline or fixed perspectives. To address the above limitations, we propose an image stitching network consisting of three modules: multistage keypoint matching module, DFAST module and multistage image reconstruction module. First, we use a multistage keypoint matching module to align the reference image and the target image, and obtain deep homography estimates between reference and target images at different scales of features. After that, the DFAST module is designed to stitch images of arbitrary views and generate stitched feature maps at different scales. Finally, the multistage reconstruction network is used to reconstruct and optimize the stitched feature maps from feature level to pixel level and fuse stitched images of different scales to generate finer texture details. Experiments results show that our method surpasses previous methods including state-of-the-art traditional and CNN-based methods.
Semantic segmentation is vital for computers to process image scene parsing. Semantic segmentation requires the output of high-resolution segmentation results that classify each pixel in an image, and high-resolution feature maps consume significant computational costs in deep learning networks. To trade-off the accuracy and speed of semantic segmentation models, researchers have proposed various different approaches to extract features by semantic segmentation backbone networks. In this paper, we propose a novel semantic-guided backbone network (SG2Path) based on a multi-branch backbone architecture, in which we design a semantic-guided upsampling module (SGUM) to better fuse high-resolution spatial features with low-resolution semantic features in different branches, which effectively solves the semantic misalignment problem between feature maps of different resolutions. Experiments on the Cityscapes dataset and the Visualization analysis of the model prove the advantages of our model in the semantic segmentation task and the significant application potential.
In recent years, forward-scan sonar is widely applied to the underwater inspection, which is not subject to the influence
of light and turbidity. For expanding the monitoring scope, the image sonar is generally mounted on the pan-tilt platform
of a ROV (Remotely Operated Vehicle) or survey boat. However, there are still some problems such as: 1) The field-of-view
is narrow, i.e. the horizontal view angle of DIDSON (Dual-frequency identification sonar) is 29°; 2) The dynamic
change of a ROV or survey boat by the water disturbances will cause the target to escape from the sonar image easily; 3)
The image sonar is fixed on the pan-tilt platform, and its position and posture are unceasingly changed. As a result of
these problems, the obtained images may be distorted and not on the same plane. To solve the above problems, stability
augmentation of pan-tilt platform based on the principle of bionic eye movements and a mosaic method of sonar images
are presented. According to the principle of the vestibule-ocular reflex, an active compensation control system of the
mechanical pan-tilt platform is developed. It can compensate the sonar image instability resulting from attitude variation
of a ROV or survey boat during operation. Applying multi-sensor fusion technology can rectify the sonar images with
different position and posture to be on a single geodetic coordinate frame for image matching. Finally, sonar images can
be mosaic. A stable large-scale sonar image can be obtained. The experimental results validate that the presented method
is valid.
KEYWORDS: Robotics, Control systems, Control systems design, Sensors, Telecommunications, Actuators, Navigation systems, Human-machine interfaces, Signal processing, Aerodynamics
Robotic blimps present an enormous potential for applications in low-speed and low-altitude exploration, surveillance, and monitoring, as well as telecommunication relay platforms. To make our lighter-than-air platform a robotic blimp with significant levels of autonomy, the decoupled longitude and latitude dynamic model are developed, and the hardware and software of the flight control system are designed. The onboard hardware consists of blimp state observer, actuators, MCU, etc. The software functions include signals processing, data filtering and fault tolerance, ground command execution, etc. Based on decoupled dynamic model, the control system architecture is presented, and navigation strategy for waypoint flight problem is discussed. The paper gives results of a flight experiment using the designed flight control system, and the results manifests that the system is applicable and initial machine intelligence of robotic blimp is achieved.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.