Pub Date : 2022-07-17DOI: 10.1109/RCAR54675.2022.9872182
Xi Wang, Weilin Chen, Jiaole Wang, Shuang Song
Accurate positioning and efficient movement are essential for magnetic capsule endoscopy, which has attracted more and more attention in recent years. However, moving in the desired trajectory often conflicts with precise positioning, as magnetic localization is only feasible in a small area near the sensors. In this paper, we proposed a closed-loop magnetic capsule robot actuation system, which can accomplish localization and actuation simultaneously on a large scale in the fluid environment of the human body. To achieve large-scale detection, electromagnetic coil and sensor array are fixed together on a 3-axis screw mobile platform. The distribution of magnetic field is analyzed with magnetic dipole model and rectangular electromagnetic coil model. Levenberg-Marquardt algorithm has been employed to estimate the position of the capsule robot by subtracting the actuation magnetic field. PI closed-loop controller with localization of the robot as feedback is applied in the system. Although the response speed of the system with the PI controller is not fast, it could perform well in stability, which is expected when the capsule is moving inside the human body. Two specific path following experiments were carried out to verify the performance of simultaneous localization and movement on a large scale. Results showed that the proposed system and method could work well.
{"title":"Closed-loop Electromagnetic Actuation System for Magnetic Capsule Robot In a Large Scale","authors":"Xi Wang, Weilin Chen, Jiaole Wang, Shuang Song","doi":"10.1109/RCAR54675.2022.9872182","DOIUrl":"https://doi.org/10.1109/RCAR54675.2022.9872182","url":null,"abstract":"Accurate positioning and efficient movement are essential for magnetic capsule endoscopy, which has attracted more and more attention in recent years. However, moving in the desired trajectory often conflicts with precise positioning, as magnetic localization is only feasible in a small area near the sensors. In this paper, we proposed a closed-loop magnetic capsule robot actuation system, which can accomplish localization and actuation simultaneously on a large scale in the fluid environment of the human body. To achieve large-scale detection, electromagnetic coil and sensor array are fixed together on a 3-axis screw mobile platform. The distribution of magnetic field is analyzed with magnetic dipole model and rectangular electromagnetic coil model. Levenberg-Marquardt algorithm has been employed to estimate the position of the capsule robot by subtracting the actuation magnetic field. PI closed-loop controller with localization of the robot as feedback is applied in the system. Although the response speed of the system with the PI controller is not fast, it could perform well in stability, which is expected when the capsule is moving inside the human body. Two specific path following experiments were carried out to verify the performance of simultaneous localization and movement on a large scale. Results showed that the proposed system and method could work well.","PeriodicalId":304963,"journal":{"name":"2022 IEEE International Conference on Real-time Computing and Robotics (RCAR)","volume":"33 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115952495","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2022-07-17DOI: 10.1109/RCAR54675.2022.9872236
Mengmeng Wang, Cai-zhi Fan, Chao Song
Small video satellites are capable of conducting real-time continuous observation of space targets through attitude control and have broad application prospects. Since the traditional method of tracking based on location information needs the priori location information of the known target, effective tracking observation cannot be accomplished for non-cooperative targets. In this paper, we are going to design a visual tracking attitude control method for spatial targets based on image information, which can perform autonomous tracking observation for both cooperative and non-cooperative targets. Firstly, based on the principle of perspective projection, the internal and external parameter model of the camera is derived, and the conversion relationship between the inertial coordinate system and the pixel coordinate system of the on-board camera is established. Then the attitude dynamical model and kinematical model of the rigid satellite are given. The desired attitude and desired angular velocity of the small video satellite are derived based on the deviation information of the location coordinates of the target in the image plane projection point from the desired coordinates. Using the attitude error and angular velocity error as the control feedback quantity, the space target tracking PD controller is designed. The global stability of the closed-loop control system is proved using Barbalat theorem. The simulation results show that the proposed control method is effective for the visual tracking attitude control of space targets.
{"title":"Image-Based Visual Tracking Attitude Control Research on Small Video Satellites for Space Targets","authors":"Mengmeng Wang, Cai-zhi Fan, Chao Song","doi":"10.1109/RCAR54675.2022.9872236","DOIUrl":"https://doi.org/10.1109/RCAR54675.2022.9872236","url":null,"abstract":"Small video satellites are capable of conducting real-time continuous observation of space targets through attitude control and have broad application prospects. Since the traditional method of tracking based on location information needs the priori location information of the known target, effective tracking observation cannot be accomplished for non-cooperative targets. In this paper, we are going to design a visual tracking attitude control method for spatial targets based on image information, which can perform autonomous tracking observation for both cooperative and non-cooperative targets. Firstly, based on the principle of perspective projection, the internal and external parameter model of the camera is derived, and the conversion relationship between the inertial coordinate system and the pixel coordinate system of the on-board camera is established. Then the attitude dynamical model and kinematical model of the rigid satellite are given. The desired attitude and desired angular velocity of the small video satellite are derived based on the deviation information of the location coordinates of the target in the image plane projection point from the desired coordinates. Using the attitude error and angular velocity error as the control feedback quantity, the space target tracking PD controller is designed. The global stability of the closed-loop control system is proved using Barbalat theorem. The simulation results show that the proposed control method is effective for the visual tracking attitude control of space targets.","PeriodicalId":304963,"journal":{"name":"2022 IEEE International Conference on Real-time Computing and Robotics (RCAR)","volume":"132 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116755465","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The manipulators usually need to contact with the environment when executing the tasks. Maintaining the stability of the contact force between the manipulator end-effector and the environment is very crucial. However, constant admittance control method cannot maintain the stability of dynamic force tracking if the environment is uncalibrated. A variable admittance control algorithm based on reinforcement learning is proposed, which adjusts the damping parameter of admittance control through reinforcement learning agent. Through the simulation experiments, it is found that this method can maintain the stability of dynamic contact force tracking on a sloped surface and a sine surface when an estimation error of the environmental position exists. Compared with the traditional admittance control with constant coefficients, the adaptive admittance control algorithm performs better.
{"title":"Variable Admittance Control for Robotic Contact Force Tracking in Dynamic Environment Based on Reinforcement Learning","authors":"Yufei Zhou, Tianyu Liu, Jingkai Cui, Yanhui Li, Mingchao Zhu","doi":"10.1109/RCAR54675.2022.9872292","DOIUrl":"https://doi.org/10.1109/RCAR54675.2022.9872292","url":null,"abstract":"The manipulators usually need to contact with the environment when executing the tasks. Maintaining the stability of the contact force between the manipulator end-effector and the environment is very crucial. However, constant admittance control method cannot maintain the stability of dynamic force tracking if the environment is uncalibrated. A variable admittance control algorithm based on reinforcement learning is proposed, which adjusts the damping parameter of admittance control through reinforcement learning agent. Through the simulation experiments, it is found that this method can maintain the stability of dynamic contact force tracking on a sloped surface and a sine surface when an estimation error of the environmental position exists. Compared with the traditional admittance control with constant coefficients, the adaptive admittance control algorithm performs better.","PeriodicalId":304963,"journal":{"name":"2022 IEEE International Conference on Real-time Computing and Robotics (RCAR)","volume":"19 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131757303","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2022-07-17DOI: 10.1109/RCAR54675.2022.9872242
Jun Yu Li, Yunshuang Zhang, Shuai Zhao, Chen Chao, Zhibin Du
Because of the safety risks caused by functional insufficiencies and performance limitations for automated vehicle, the Safety of The Intended Function (SOTIF) of the Lane Keep Assistance (LKA) system is studied based on the System-Theoretic Process Analysis (SPTA). The interaction of LKA system control model with driver, steering system, data acquisition system and other external environment is established. Based on the model, 7 kinds of Unsafe Control Actions (UCA) are identified, and the vehicle-level safety constrains are proposed. 20 triggering conditions are identified from the perspectives of functional insufficiency and misuse. Taking the severity and controllability as the evaluation indexes, the risk assessment of each trigger condition is carried out, and the improvement measures are put forward. This study comprehensively reveals the way to realize the intended functional safety of LKA, and lays a foundation for the formulation of the control strategy of autonomous vehicles.
{"title":"A Research on SOTIF of LKA based on STPA*","authors":"Jun Yu Li, Yunshuang Zhang, Shuai Zhao, Chen Chao, Zhibin Du","doi":"10.1109/RCAR54675.2022.9872242","DOIUrl":"https://doi.org/10.1109/RCAR54675.2022.9872242","url":null,"abstract":"Because of the safety risks caused by functional insufficiencies and performance limitations for automated vehicle, the Safety of The Intended Function (SOTIF) of the Lane Keep Assistance (LKA) system is studied based on the System-Theoretic Process Analysis (SPTA). The interaction of LKA system control model with driver, steering system, data acquisition system and other external environment is established. Based on the model, 7 kinds of Unsafe Control Actions (UCA) are identified, and the vehicle-level safety constrains are proposed. 20 triggering conditions are identified from the perspectives of functional insufficiency and misuse. Taking the severity and controllability as the evaluation indexes, the risk assessment of each trigger condition is carried out, and the improvement measures are put forward. This study comprehensively reveals the way to realize the intended functional safety of LKA, and lays a foundation for the formulation of the control strategy of autonomous vehicles.","PeriodicalId":304963,"journal":{"name":"2022 IEEE International Conference on Real-time Computing and Robotics (RCAR)","volume":"27 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115136261","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2022-07-17DOI: 10.1109/RCAR54675.2022.9872286
Meiyuan Zou, Jiajie Yu, Bo Lu, Wenzheng Chi, Lining Sun
As a common multi-functional engineering equipment, excavators are widely used in civil construction, coal mining, power engineering, etc. The excellent performance of the excavator not only greatly improves the work efficiency during the construction process, but also effectively saves labor costs. However, due to the complexity of the working environment of the excavator and the blind area of the excavator itself, the driver cannot make timely judgments on the surrounding environment, which may cause potential threats to pedestrians. In response to such problems, this paper proposes a multi-sensor fusion detection method applied to excavators to provide vision assistance for excavator drivers, thereby reducing the risk of pedestrian casualties. Based on the results of the joint calibration, the transformation relationship between the camera and lidar coordinate systems is determined. Combining the detection results of the pedestrian detection algorithm YOLO-v5 and the segmented image information, the position of the pedestrian in the image can be inversely mapped to the 3D point clouds via the matrix transformation, which can accurately display the position of the pedestrian in the point clouds, consequently making up for the lack of depth information in the image. The experimental results show that our method can effectively extract the location information of pedestrians from the complex background environment and realize timely pedestrian alarm.
{"title":"Active Pedestrian Detection for Excavator Robots based on Multi-Sensor Fusion","authors":"Meiyuan Zou, Jiajie Yu, Bo Lu, Wenzheng Chi, Lining Sun","doi":"10.1109/RCAR54675.2022.9872286","DOIUrl":"https://doi.org/10.1109/RCAR54675.2022.9872286","url":null,"abstract":"As a common multi-functional engineering equipment, excavators are widely used in civil construction, coal mining, power engineering, etc. The excellent performance of the excavator not only greatly improves the work efficiency during the construction process, but also effectively saves labor costs. However, due to the complexity of the working environment of the excavator and the blind area of the excavator itself, the driver cannot make timely judgments on the surrounding environment, which may cause potential threats to pedestrians. In response to such problems, this paper proposes a multi-sensor fusion detection method applied to excavators to provide vision assistance for excavator drivers, thereby reducing the risk of pedestrian casualties. Based on the results of the joint calibration, the transformation relationship between the camera and lidar coordinate systems is determined. Combining the detection results of the pedestrian detection algorithm YOLO-v5 and the segmented image information, the position of the pedestrian in the image can be inversely mapped to the 3D point clouds via the matrix transformation, which can accurately display the position of the pedestrian in the point clouds, consequently making up for the lack of depth information in the image. The experimental results show that our method can effectively extract the location information of pedestrians from the complex background environment and realize timely pedestrian alarm.","PeriodicalId":304963,"journal":{"name":"2022 IEEE International Conference on Real-time Computing and Robotics (RCAR)","volume":"66 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131462165","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2022-07-17DOI: 10.1109/RCAR54675.2022.9872180
Wen Fu, Yanjie Li, Zhaohui Ye, Qi Liu
On the basis of environmental information processed by the sensing module, the decision module in automatic driving integrates environmental and vehicle information to make the autonomous vehicle produce safe and reasonable driving behavior. Considering the complexity and variability of the driving environment of autonomous vehicles, researchers have begun to apply deep reinforcement learning (DRL) in the study of autonomous driving control strategies in recent years. In this paper, we apply an algorithm framework combining multimodal transformer and DRL to solve the autonomous driving decision problem in complex scenarios. We use ResNet and transformer to extract the features of LiDAR point cloud and image. We use Deep Deterministic Policy Gradient (DDPG) algorithm to complete the subsequent autonomous driving decision-making task. And we use information bottleneck to improve the sampling efficiency of RL. We use CARLA simulator to evaluate our approach. The results show that our approach allows the agent to learn better driving strategies.
{"title":"Decision Making for Autonomous Driving Via Multimodal Transformer and Deep Reinforcement Learning*","authors":"Wen Fu, Yanjie Li, Zhaohui Ye, Qi Liu","doi":"10.1109/RCAR54675.2022.9872180","DOIUrl":"https://doi.org/10.1109/RCAR54675.2022.9872180","url":null,"abstract":"On the basis of environmental information processed by the sensing module, the decision module in automatic driving integrates environmental and vehicle information to make the autonomous vehicle produce safe and reasonable driving behavior. Considering the complexity and variability of the driving environment of autonomous vehicles, researchers have begun to apply deep reinforcement learning (DRL) in the study of autonomous driving control strategies in recent years. In this paper, we apply an algorithm framework combining multimodal transformer and DRL to solve the autonomous driving decision problem in complex scenarios. We use ResNet and transformer to extract the features of LiDAR point cloud and image. We use Deep Deterministic Policy Gradient (DDPG) algorithm to complete the subsequent autonomous driving decision-making task. And we use information bottleneck to improve the sampling efficiency of RL. We use CARLA simulator to evaluate our approach. The results show that our approach allows the agent to learn better driving strategies.","PeriodicalId":304963,"journal":{"name":"2022 IEEE International Conference on Real-time Computing and Robotics (RCAR)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131687619","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2022-07-17DOI: 10.1109/RCAR54675.2022.9872231
Yangyan Deng, Ding Yuan, Hong Zhang
With the development of deep learning, multi-view stereo has achieved significant progress recently. Due to the expensive three-dimension supervision, self-supervised methods have more potential. In this work, a novel two-stage self-supervised learning framework for multi-view stereo is proposed to overcome photometric dependency and the effect of foreshortening. On considering that accurate depth hypothesis always plays an important role in estimating depth information. Therefore, this work concentrates on designing an adaptive depth sampling module based on neighboring spatial patches propagation, which is determined by the normal maps. From this point of view, a two-stage process is carried out in this work. In detail, the coarse initial depth maps and normal maps are obtained in the first stage, and then the network in the second stage refines the depth sampling module by taking the influence of foreshortening into account. Furthermore, the loss functions are developed including feature-metric consistency to overcome the photometric inconsistency caused by lighting variation. Moreover, the consistency between depth maps and normal maps is also employed in the loss functions. To evaluate the effectiveness of our proposed two-stage framework, the experiments are carried out on the DTU datasets. The experimental results demonstrate that our self-supervised learning framework has excellent performance compared to the baseline methods.
{"title":"Two-stage Self-supervised MVS Network using Adaptive Depth Sampling","authors":"Yangyan Deng, Ding Yuan, Hong Zhang","doi":"10.1109/RCAR54675.2022.9872231","DOIUrl":"https://doi.org/10.1109/RCAR54675.2022.9872231","url":null,"abstract":"With the development of deep learning, multi-view stereo has achieved significant progress recently. Due to the expensive three-dimension supervision, self-supervised methods have more potential. In this work, a novel two-stage self-supervised learning framework for multi-view stereo is proposed to overcome photometric dependency and the effect of foreshortening. On considering that accurate depth hypothesis always plays an important role in estimating depth information. Therefore, this work concentrates on designing an adaptive depth sampling module based on neighboring spatial patches propagation, which is determined by the normal maps. From this point of view, a two-stage process is carried out in this work. In detail, the coarse initial depth maps and normal maps are obtained in the first stage, and then the network in the second stage refines the depth sampling module by taking the influence of foreshortening into account. Furthermore, the loss functions are developed including feature-metric consistency to overcome the photometric inconsistency caused by lighting variation. Moreover, the consistency between depth maps and normal maps is also employed in the loss functions. To evaluate the effectiveness of our proposed two-stage framework, the experiments are carried out on the DTU datasets. The experimental results demonstrate that our self-supervised learning framework has excellent performance compared to the baseline methods.","PeriodicalId":304963,"journal":{"name":"2022 IEEE International Conference on Real-time Computing and Robotics (RCAR)","volume":"9 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116602419","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
In order to improve the prediction accuracy of the thermal error model of CNC machine tools based on time series and reduce the time of model parameter identification, a time series thermal error modeling method based on intelligent optimization (GA-ARMA) was proposed. Using the reciprocal of the residual between the actual value and the predicted value as the genetic algorithm (GA) individual fitness value function, select the best individual obtained by evolution for several generations as the parameter of the ARMA model, quickly identify the parameters of the ARMA model, and establish the GA-ARMA spindle axial thermal error model. Through experiments to compare the prediction effects of the time series thermal error model based on intelligent optimization and the time series thermal error model, taking a certain type of three-axis CNC machine tool as the object, the prediction and comparison are carried out under different working conditions. The experimental results show that the model prediction average residual error reaches 1.28 $mu$m, and the modeling efficiency is improved by 544%.
{"title":"Research on Axial Thermal Error Modeling Method of CNC Machine Tool Spindle Based on GA-ARMA*","authors":"Weicheng Lin, Ling Yin, Fei Zhang, Zewei He, Yu Chen, Wenhao Li, Yeming Song","doi":"10.1109/RCAR54675.2022.9872211","DOIUrl":"https://doi.org/10.1109/RCAR54675.2022.9872211","url":null,"abstract":"In order to improve the prediction accuracy of the thermal error model of CNC machine tools based on time series and reduce the time of model parameter identification, a time series thermal error modeling method based on intelligent optimization (GA-ARMA) was proposed. Using the reciprocal of the residual between the actual value and the predicted value as the genetic algorithm (GA) individual fitness value function, select the best individual obtained by evolution for several generations as the parameter of the ARMA model, quickly identify the parameters of the ARMA model, and establish the GA-ARMA spindle axial thermal error model. Through experiments to compare the prediction effects of the time series thermal error model based on intelligent optimization and the time series thermal error model, taking a certain type of three-axis CNC machine tool as the object, the prediction and comparison are carried out under different working conditions. The experimental results show that the model prediction average residual error reaches 1.28 $mu$m, and the modeling efficiency is improved by 544%.","PeriodicalId":304963,"journal":{"name":"2022 IEEE International Conference on Real-time Computing and Robotics (RCAR)","volume":"54 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116754374","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2022-07-17DOI: 10.1109/RCAR54675.2022.9872179
Tongtong Wang, R. Skulstad, Motoyasu Kanazawa, Guoyuan Li, V. Æsøy, Houxiang Zhang
Accurate ship motion predictions play a vital role in supporting the decision-making process onboard. Generally, the ship dynamics are described by either a deterministic model derived from hydrodynamic principles or a black-box model learned from the observations. However, there are always cases in real life where the physics information is insufficient to develop a complete model, and the data quantity is also limited so that a data-driven model is away from expectation. For this obstacle, we propose a physics-data cooperative modeling approach based on a rough ship numerical model and a few operational data to enhance the model quality. The prior knowledge leveraged by the ship’s numerical model is integrated into the neural network as informative inputs, and the informed neural network calibrates the bias between model outcomes and actual states in principle. The proposed approach is validated in the real docking operation of a research vessel. Comparisons with both the purely hydrodynamic model and the data-driven model without physics informed are conducted. The results convinced that the physicsdata hybrid way yields a more accurate model with relaxed data requirements and less learning consumption.
{"title":"Physics-informed Data-driven Approach for Ship Docking Prediction","authors":"Tongtong Wang, R. Skulstad, Motoyasu Kanazawa, Guoyuan Li, V. Æsøy, Houxiang Zhang","doi":"10.1109/RCAR54675.2022.9872179","DOIUrl":"https://doi.org/10.1109/RCAR54675.2022.9872179","url":null,"abstract":"Accurate ship motion predictions play a vital role in supporting the decision-making process onboard. Generally, the ship dynamics are described by either a deterministic model derived from hydrodynamic principles or a black-box model learned from the observations. However, there are always cases in real life where the physics information is insufficient to develop a complete model, and the data quantity is also limited so that a data-driven model is away from expectation. For this obstacle, we propose a physics-data cooperative modeling approach based on a rough ship numerical model and a few operational data to enhance the model quality. The prior knowledge leveraged by the ship’s numerical model is integrated into the neural network as informative inputs, and the informed neural network calibrates the bias between model outcomes and actual states in principle. The proposed approach is validated in the real docking operation of a research vessel. Comparisons with both the purely hydrodynamic model and the data-driven model without physics informed are conducted. The results convinced that the physicsdata hybrid way yields a more accurate model with relaxed data requirements and less learning consumption.","PeriodicalId":304963,"journal":{"name":"2022 IEEE International Conference on Real-time Computing and Robotics (RCAR)","volume":"108 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123536424","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2022-07-17DOI: 10.1109/RCAR54675.2022.9872237
Qianyi Zhang, Dingye Yang, Lei Zhou, Zhengxi Hu, Jingtai Liu
Focused on the balance among safety, length, and smoothness, this paper proposes a novel model to train an agent with deep reinforcement learning to optimize trajectory in complex environments. Inspired by the human habit that first finds the shortest trajectory and then slightly optimizes safety and smoothness, State is initialized as a radical trajectory combined with local obstacle distribution. Action adjusts dangerous waypoints jointly. Reward penalizes length increase based on local smoothness change. Episode is early terminated to divide the whole problem into smaller ones, while reward assembles them back with a large amount of training data. This allows the agent to be trained locally and work globally to accelerate convergence. Performances in various scenarios demonstrate our method’s ability to balance safety, length, and smoothness. With the Markov property of the problem and our newly discovered mathematical property of B-spline, it adjusts waypoints under sub-grid map and can be generalized stably in various maps with dense obstacles.
{"title":"Trajectory Optimization on Safety, Length and Smoothness in Complex Environments with A Locally Trained and Globally Working Agent","authors":"Qianyi Zhang, Dingye Yang, Lei Zhou, Zhengxi Hu, Jingtai Liu","doi":"10.1109/RCAR54675.2022.9872237","DOIUrl":"https://doi.org/10.1109/RCAR54675.2022.9872237","url":null,"abstract":"Focused on the balance among safety, length, and smoothness, this paper proposes a novel model to train an agent with deep reinforcement learning to optimize trajectory in complex environments. Inspired by the human habit that first finds the shortest trajectory and then slightly optimizes safety and smoothness, State is initialized as a radical trajectory combined with local obstacle distribution. Action adjusts dangerous waypoints jointly. Reward penalizes length increase based on local smoothness change. Episode is early terminated to divide the whole problem into smaller ones, while reward assembles them back with a large amount of training data. This allows the agent to be trained locally and work globally to accelerate convergence. Performances in various scenarios demonstrate our method’s ability to balance safety, length, and smoothness. With the Markov property of the problem and our newly discovered mathematical property of B-spline, it adjusts waypoints under sub-grid map and can be generalized stably in various maps with dense obstacles.","PeriodicalId":304963,"journal":{"name":"2022 IEEE International Conference on Real-time Computing and Robotics (RCAR)","volume":"55 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121808840","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：