Research On Robot Deep Learning Inference Technology In Mobile Ad-hoc Cloud

In recent years,more and more robotic problems have begun to explore solutions based on deep learning methods.However,the improvement of the ability of deep neural network is often accompanied with the increase of the model size and complexity,which needs to consume plenty of resources at runtime.As a kind of special computing equipment,the resources of robots are often limited because of the physical design constraints.The contradiction between them has become a bottleneck problem that restricts the improvement of robot autonomy and intelligence.In the field of distributed systems,a feasible solution to this problem is to use Ad Hoc Cloud,that is,multiple robots form virtual organizations dynamically by using cloud computing technology in a selforganizing and peer-to-peer way,sharing the required computing resources,and jointly completing the computing-intensive tasks.The architecture,mechanism and implementation methods of Ad Hoc Cloud have already been studied in the fields of mobile computing.However,the existing works are mainly limited to the support of ordinary mobile applications,rather than the support of robot "intelligence".This leads to two problems:(1)Lack of computational offloading methods and scheduling mechanisms for inference tasks based on deep learning.Deep learning models have significant computing and data intensive characteristics,and maintain the characteristic of progressive by layers in calculation.At present,there is no relevant research on how such models can be reasonably organized and scheduled in the Ad Hoc Cloud so that the performance of computing nodes cannot be made best of use.(2)Lack of research on real-time maintenance technology for inference tasks based on deep learning.Traditional implementations of Ad Hoc Cloud don't focus on real-time performance which is instead a key consideration for robotic tasks(such as autonomous obstacle avoidance and unmanned aerial vehicle tracking).According to the two challenges mentioned above,this research carried out works from three aspects as follows:(1)Constructing the model of behavior feature and resource requirements of deep learning inference algorithm of the robot.The analysis of the behavior feature and resource requirements of typical deep learning inference algorithms is a prerequisite for properly scheduling deep learning computing tasks in the Ad Hoc Cloud.This research topic utilizes classical robotic platform to conduct empirical research.We collected and analyzed the performance characteristics and resource consumption data at run time of the deep learning model inference process,and obtained the knowledge of the time and space complexity of computing,resource requirements,and execution bottlenecks of each layer in the typical deep learning models.(2)We proposed a model partition based approach for deep learning inference computing offloading in the Ad Hoc Cloud.In order to solve the contradiction between the resource requirements for deep learning inference and the limitation of robot resources when the data center is not available,this research adopts model partition-based approach to perform offloading of the computationally intensive part of deep learning model inference,which decreases the execution threshold and implements distributed deep learning model inference based on local robots collaboration.(3)We proposed a timeliness adaptation method for deep learning inference in homogeneous Ad Hoc Cloud.For the time-sensitive tasks based on deep learning,on the basis of the Ad Hoc Cloud execution environment composing homogeneous nodes,this research proposed a model parallel and data parallel optimization mechanism for deep learning inference and achieved the timeliness-oriented collaboratively deep learning model inference supported by homogeneous robots.