KR20190041840A - Controlling mobile robot based on asynchronous target classification - Google Patents

Abstract

Provided is technology controlling a mobile robot based on asynchronous target classification. According to embodiments of the present invention, unlike an end-to-end robot software architecture, the technology may effectively integrate large-scale deep neural networks into a mobile robot requiring high bandwidth control using an asynchronous deep classification network (ADCN), which may introduce deep learning technology as one component constituting the entire software of the mobile robot.

Description

[0001] CONTROLLING MOBILE ROBOT BASED ON ASYNCHRONOUS TARGET CLASSIFICATION [0002]

The following description relates to a mobile robot control technology based on asynchronous target classification, and more particularly to an asynchronous Deep Classification Network (ADCN) that can classify targets asynchronously while using deep learning techniques And a computer program stored in a computer-readable recording medium for causing a computer to execute a method of classifying a target, the recording medium being associated with the computer.

Due to the breakthrough of deep learning technology, some prior arts have suggested the possibility of applying deep learning to the mobile robot field. For example, in the following paper 1, a mobile robot structure with end-to-end motion planning for obstacle avoidance and path search using a deep neural network is proposed. In the following paper 2, We proposed a training methodology for end - to - end visual navigation and deep neural networks using neural networks.

- Papers 1: M. Pfeiffer, M. Schaeuble, J. Nieto, R. Siegwart, and C. Cadena, "From perception to decision: A data-driven approach to autonomous ground robots, IEEE International Conference on Robotics and Automation (ICRA), 2017.

- Papers 2: Y. Zhu, R. Mottaghi, E. Kolve, J Lim, JJ, A. Gupta, L. Fei-Fei, and A. Farhadi, "Target-driven visual navigation in indoor scenes using deep reinforcement learning, " IEEE International Conference on Robotics and Automation (ICRA), 2017.

Although these techniques are new, they show the following limitations from a traditional robot perspective.

- Low compatibility with previously developed robot software and hardware components.

- It is difficult to operate in a small mobile robot because it requires large computing performance.

- It is difficult to apply application-specific training methods of reinforcement learning to other robot platforms.

With regard to the visual recognition of the robot, the deep recognition based visual recognition method provides excellent recognition performance compared to the conventional feature-based recognition method. In addition, recently, these deep learning based visual recognition methods have been extended to other areas of conventional computer vision, such as image segmentation. Considering performance and scalability, it is very useful to use a deep learning - based approach to various recognition tasks such as visual recognition of mobile robots.

However, unlike in other applications, the limited computing performance of mobile robots is a major obstacle to using a deep learning based approach in the visual recognition system of mobile robots. Most high performance algorithms used in deep image recognition in the prior art require a large amount of computation due to the highly stacked network structure. Also, considering the requirement for use of advanced hardware according to the amount of computation, embedding such a prior art network in a small mobile robot greatly reduces the throughput of the visual recognition system of the mobile robot.

This throughput problem further degrades system performance when applying recently studied end-to-end models to mobile robots. Since the end-to-end model connects the visual recognition system of the mobile robot and the control system of the mobile robot through a single network, the bandwidth of the visual recognition calculation is remarkably low, thereby deteriorating the overall throughput of the system.

As a solution to this problem, there may be a method of using an external server with a high-level GPU (Graphic Processing Unit) for visual recognition network and a method of reducing and limiting the number of layers of the neural network to increase the throughput of the network . However, in the former case, the waiting time between the external server and the mobile robot may have a great influence on the throughput of the mobile robot as a result of transmitting a plurality of images to the external server. For example, the GPU of the external server may receive a whole image input through the camera of the mobile robot in order to supply the input of the algorithm, so that the communication overhead may be fatal. Also, in the latter case, there is a problem that the pruning network structure reduces the recognition performance of the network. For example, a structurally optimized neural network with increased network throughput by reducing and limiting the number of layers in a neural network is known to be less accurate than a heavy network of other prior art techniques.

There is also a reinforcement learning technique that uses an image as the input state of the network, and this conventional reinforcement learning technique allows the neural network to learn the nuances of the situation in the input image. Due to the ability to understand the complex context of an image, deep reinforcement learning can be applied to robot control systems such as path planning or robotic arm manipulation. Unlike the past approach, where only low-dimensional data such as proximity information or the speed of the robot could be used, a recent approach utilizes a raw camera image as input. Many robot control algorithms with an end-to-end concept have been proposed, for example, as it has become possible to link raw image of a camera to a motor control signal in a single neural network, End visuomotor network can be applied to real robot control with fine tuning in the robot platform.

Even though such an end-to-end controller scheme has its own superiority, this method has a problem in that compatibility is generally limited as described above. If an end-to-end controller network is learned and embedded in a given robot platform, it is very difficult to apply this network to other sensors and dynamics of other robots. Also, it takes a lot of time and effort to learn reinforcement learning network with real robot hardware. Therefore, it is necessary to construct a simulation environment for learning reinforcement learning network. However, there is a problem that a very complicated process such as 3D space reconfiguration is required to develop a realistic learning simulator.

Unlike the end-to-end robot software architecture, in order to effectively integrate large-scale deep neural networks into mobile robots requiring high bandwidth control, Deep Learning technology is introduced as a component of the overall software of mobile robots Asynchronous Deep Classification Network (ADCN), which can be used to provide a wide range of services.

Gaming Reinforcement Learning-Based Motion Planner (GRL-Planner) for mobile robots is provided to reduce learning complexity and improve algorithm scalability.

Obtaining image information about a position of a target and a region of interest corresponding to the target in the image through image processing of an image inputted through a camera included in the mobile robot; Mapping the position of the target on a two-dimensional plane to generate an abstraction map; Storing image information for a region of interest corresponding to the target in the image, corresponding to a position of the target mapped to the abstraction map; Processing image information for a region of interest stored in the abstraction map through an asynchronous Deep Classification Network to classify the target asynchronously; And updating the abstraction map with information about asynchronously classified targets through the asynchronous deep classification network.

There is provided a computer program stored in a computer-readable recording medium for causing a computer to execute the target classification method in combination with a computer.

There is provided a computer-readable recording medium having recorded thereon a program for causing a computer to execute the target classification method.

A computer apparatus for controlling a mobile robot, comprising: a memory for storing a computer-readable instruction; And at least one processor configured to execute the command, wherein the at least one processor is operable to determine, based on the position of the target through image processing on the image input through the camera included in the mobile robot, Acquiring image information for a region of interest corresponding to the target, mapping the position of the target on a two-dimensional plane to generate an abstraction map, and generating image information for a region of interest corresponding to the target in the image, And asynchronously classifying the target asynchronously by processing image information on a region of interest stored in the abstraction map through an asynchronous Deep Classification Network, Classify asynchronously through deep classifier networks Information about the target to provide a computer device, characterized in that updating the map abstraction.

1. A mobile robot, comprising: a memory for storing instructions readable by a computer; At least one processor configured to execute the instruction; And a camera for receiving an image, wherein the at least one processor is operable to perform image processing on an image input through the camera to obtain image information on a position of the target and a region of interest corresponding to the target in the image Generating an abstraction map by mapping the position of the target on a two-dimensional plane, and generating image information for a region of interest corresponding to the target in the image, corresponding to a position of the target mapped to the abstraction map Processing the image information for a region of interest stored in the abstraction map through an asynchronous Deep Classification Network to classify the targets asynchronously, and asynchronously classifying the objects asynchronously through the asynchronous Deep Classification Network Information about the target is updated in the abstraction map The mobile robot includes:

Unlike the end-to-end robot software architecture, the Asynchronous Deep Classification Network (ADCN), which can introduce deep learning technology as a component of the overall software of mobile robots, It is possible to effectively integrate a large-scale deep neural network into a mobile robot requiring control.

The Gaming Reinforcement Learning-Based Motion Planner (GRL-Planner) for mobile robots can reduce learning complexity and improve algorithm scalability.

1 is a schematic diagram of a high bandwidth framework for an asynchronous deep classifier network, in accordance with an embodiment of the present invention.
2 is a diagram illustrating an example of a process of extracting data and generating an abstraction map according to an embodiment of the present invention.
FIG. 3 is a diagram illustrating an example of an overall process of how a reinforcement learning based motion planner operates in a mobile robot platform, in an embodiment of the present invention.
4 is a diagram illustrating an example of a mobile robot according to an embodiment of the present invention.
5 is a view showing an example of controlling the inner radius of the nozzle in an embodiment of the present invention.
6 is a diagram illustrating examples of an object recognition screen and an abstract map screen according to an exemplary embodiment of the present invention.
FIG. 7 is a diagram illustrating an example of a state diagram of a multi-body tracker including four states according to an embodiment of the present invention.
8 is a diagram showing an example of a simulation environment in an embodiment of the present invention.
9 is a diagram illustrating examples of the results of a robot path plan in an embodiment of the present invention.
10 is a block diagram showing an example of the internal configuration of a mobile robot according to an embodiment of the present invention.
11 is a flowchart showing an example of a target classification method according to an embodiment of the present invention.
12 is a flowchart illustrating an example of a mobile robot control method according to an embodiment of the present invention.

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings.

The representative image generation method according to embodiments of the present invention can be performed through a computer device such as a server to be described later. At this time, a computer program according to an embodiment of the present invention can be installed and operated in the computer device, and the computer device can perform the representative image generation method according to an embodiment of the present invention, under the control of a driven computer program have. The above-described computer program may be stored in a computer-readable recording medium in combination with a computer apparatus to cause the computer to execute the representative image generating method.

1 is a schematic diagram of a high bandwidth framework for an asynchronous deep classifier network, in accordance with an embodiment of the present invention. 1 shows an essential control loop 110 and an asynchronous deep classification network (ADCN) 120 for a mobile robot. The asynchronous deep classifier network 120 may be implemented asynchronously to the essential control loop 110 to prevent system bandwidth loss while using a heavy Deep Neural Network Classifier for high precision.

The essential control loop 110 may comprise a Fast Visual Perception 111, an Abstraction Map 112, and a Motion Planner 113. Here, the abstraction map 112 may integrate the data flow between the asynchronous deep classification network 120 and the mandatory control loop 110. As the classified target information, the motion planner 113 may be provided with the abstraction map 112 as an input state.

In other words, the robot framework proposed in this embodiment can be composed of fast visual recognition 111, asynchronous deep classification network 120, and motion planner 113, and the data flow between these three components can be configured Element and an abstraction map, which is one component, as a data hub for the additional sensor. The abstraction map 112 may buffer the asynchronous data flow of the robot architecture and provide an abstract type of data input to the motion planner 113.

The fast visual recognition (111) can represent various types of simple feature-based vision processes. The result of this component, such as the rough information of the target object, can be projected in a two-dimensional plane and mapped in a simplified format in the abstraction map 112. [ At the same time, various sensor data that do not require a complex preprocessing process, such as a rider (light detection and ranging), or an IMU (Inertial Measurement Unit), can be simultaneously projected onto the abstraction map 112. The asynchronous deep classification network 120 may recognize details of the detected target object, such as an object type or shape. By implementing a multi-body tracker to compensate for the inconsistency between the slow asynchronous deep classifier network 120 and the fast visual recognition 111, a general format representing properties of the target object (not affected by dynamic movement of the robot) have.

In the case of a stand-alone robot that receives power from the outside or does not communicate with a high-performance computation device (external server), it directly applies a deep neural network, which is an algorithm for analyzing an image and determining the object is an object It is difficult to do. This is because deeper neural network based classifiers require high-performance, high-power-consuming GPUs that stand-alone robots can not accommodate. This problem can be solved by a method of transmitting an image by communication with a high-performance GPU outside the robot. However, in this case, there is a problem that a communication delay occurs and the whole recognition speed is slowed down.

Accordingly, in the present embodiment, as described above, both the fast visual recognition 111, which is a speedy and inaccurate recognition algorithm, and the deep neural network, which is a recognition algorithm with a high accuracy and a low speed of recognition, An asynchronous deep classification network 120 in which a deep neural network is constructed asynchronously may be utilized so that the performance of different speed mathematical algorithms is not interfered with each other.

At this time, the data flow of all operation algorithms can be integrated on the abstraction map 112. An object, an obstacle, and an image classification result may be stored in the abstraction map 112, and the abstraction map 112 may be simply configured to facilitate various sensor data. In addition, the abstraction map 112 can be configured in a form that can be used in the motion planner 113 for calculating the behavior of the robot as already described.

2 is a diagram illustrating an example of a process of extracting data and generating an abstraction map according to an embodiment of the present invention. Fig. 2 shows an outline of a map layer that abstracts the environment. The location of the detected object and multiple sensor data may be projected onto the abstraction map 112. For example, FIG. 2 illustrates an example in which the mobile robot 210 grasps the surrounding terrain 220 using a rider LIDAR, measures the motion of the mobile robot 210 using the IMU, And the target objects of the surrounding terrain 220 are photographed. An abstract map 112 such as the two-dimensional map 230 shown in FIG. 2 can be generated using the information obtained through various sensors. Further, the information obtained through the camera is provided to the deep classifier 240 so that the target objects can be classified more accurately in an asynchronous manner. At this time, the deep classifier 240 may be a component included in the asynchronous deep classifier network 120 and may be a deep neural network based classifier that is learned by a deep neural network. The classified information is reflected back into the abstraction map 112 so that the information of the object in accordance with the algorithms that operate asynchronously with each other can be integrated on the abstraction map 112. [

For example, the fast visual recognition 111 can recognize an object roughly in an image coming in through a camera using a simple and fast object recognition algorithm such as Blob detection or edge detection. Thereafter, the abstraction map 112 can be generated by plotting various sensing data, such as a rider, an ultrasonic sensor, etc., which need not be post-processed, on a virtual plane. At this time, a roughly recognized object can be displayed on the abstraction map 112 through the quick visual recognition 111 from the image input through the camera. At this time, the transformation map between the camera and the virtual plane may be preset and provided in advance. In addition, the abstraction map 112 may store an image entered via the camera in association with the location. For example, information about a region of interest (a cropped image including an ROI, a target object, and its surroundings) may be stored as a template of the target object. While the detected target object is being drawn in the abstraction map 112, the deep sorter 240 may asynchronously sort the images stored in the abstraction map 112 and store the classification results back into the abstraction map 112.

At this time, the motion planner 113 may receive the classified abstraction map 112 as input data, and calculate the behavior of the mobile robot 210.

While the asynchronous deep classification network 120 is categorizing objects, an algorithm for fast visual recognition 111 may display information of objects on the abstraction map 112 at a relatively faster rate. If the mobile robot 210 moves, the asynchronous deep classification network 120 will have relatively earlier information of the object as compared to the algorithm for the fast visual recognition 111, May result in inconsistencies. In order to eliminate such inconsistency, even if the mobile robot 210 or the object moves, the movement of the mobile robot 210 or the object is tracked using the multi-body tracking algorithm (the multi-body tracker described above) It is possible to prevent the target object from being recognized as another object when collating the subsequent image.

In other words, in this asynchronous deep classification network 120, the deeper neural network based classifier may include an algorithm for processing the preprocessed and extracted data from the recognition layer. After processing of the data, the deep neural network based classifier can asynchronously update the results of the deep neural network for the layer of abstraction map 112 under relatively lower bandwidth than the essential control loop 110 of the mobile robot 210 . By separating the deep learning-based classification algorithm in the essential control loop 110 of the mobile robot 210 as described above, it is possible to obtain a sufficiently high operating bandwidth to handle the throughput of a deep learning based recognition algorithm that requires heavy computing performance and delay, . ≪ / RTI >

In addition, by extracting the sensor and the time data from one single abstraction map 112, it becomes possible to develop a universal control method and / or algorithm for various mobile robots. This is because it is possible to buffer variations of the sensor or visual process on the abstraction map 112 having the same type of data representation. In addition, the abstraction map 112 may provide the same input to such a control system regardless of various forms of mobile robots.

A gaming reinforcement learning-based motion planner (GRL-planner) can be used to reduce the complexity of reinforcement learning for the control of the mobile robot 210 and improve the scalability of the algorithm. The GRL-planner according to the present embodiment can focus on three aspects: compatibility with additional sensors, convenience of simulator development, and scalability to various robot platforms.

The first key to this approach is to provide the formatted abstraction map 112 to GRL-Planner as a learning data source that enhances system compatibility. Since the previously proposed abstraction map 112 can be regarded as a common data format that incorporates various types of sensor information, this approach can provide a common foundation for the GRL-planner regardless of the sensor combination.

The abstraction map 112 provided in the input state of the reinforcement learning network can also provide convenience for the development of a simulation environment. Since these two-dimensional maps are highly abstracted with discrete data, the construction simulation environment of this map requires much less complexity than the simulator of the three-dimensional environment. Thus, learning a GRL-planner with a simple simulation environment can be established in an easier way for raw camera image-based motion planning. In addition, this method can build a simulation environment similar to a two-dimensional video game in which most of the recent reinforcement learning algorithms have shown remarkable performance. This similarity makes it easier to develop GRL-planner by stably reusing the network structure of the latest reinforcement learning algorithm.

To build a simulation environment such as a video game, the output state of the network can be designed by humans manipulating a mobile robot platform equipped with a joystick controller. In the case of a differential drive mobile robot platform, four actions (front, back, left, right) can be designed for the output state of the network, and for a holonomic mobile robot platform, Left, right) can be designed for the output state of the network. In this way, the mobility of all mobile robot platforms can be discretized and simplified to take advantage of deep reinforcement learning for path planning.

By abstracting both input state and output behavior, a network can be learned easily by applying a universally usable reinforcement learning model to a simulated abstraction map environment. The output of the network can drive the robots by mapping the probability of each state of the network output to the corresponding operation for each robot.

FIG. 3 is a diagram illustrating an example of an overall process of how a reinforcement learning based motion planner operates in a mobile robot platform, in an embodiment of the present invention. The abstraction map 112 and the joystick control signal may be used as the input state and output operation of the GRL-planar, respectively. The abstraction map 112 may be downsampled and converted to a grayscale image to reduce computational complexity before being used as input data. An example of such a process is shown in the pre-processing 310 of FIG. Depending on the platform type, the motion of the mobile robot may be mapped differently to the joystick control signal.

For example, the behavior of a typical mobile robot goes through processes such as target recognition, mapping, and behavior. Since general robot path generation methods can only search the shortest path, there is a problem that it is not suitable when there are a plurality of targets, and there is a problem that it is difficult to cope with sudden change of a map such as a failure of recognition of a target or sudden obstacle . In the embodiments of the present invention, it is possible to solve this problem by using an artificial neural network-based path generation algorithm (Deep Reinforcement Learning) instead of making a countermeasure for such problems directly by a developer. In the past, the concept of reinforcement learning existed, but due to the lack of computing power, only a small number of states could be used as inputs. Due to the development of deep learning technology, images can be used as input, which is applied in many video games. It has now been demonstrated that most simple 2D video games can be learned through reinforcement learning. In the embodiments of the present invention, movement of the mobile robot 210 can be controlled by using the already-verified reinforcement learning technique and the abstraction map 112 similar to those in the two-dimensional video games. In other words, the abstraction map 112 may be used as an input state, and reinforcement learning may be applied by abstracting the motion of the robot as a game control signal. Since the abstraction map 112 is used, it is possible to easily model the processing result of new sensor data or time data and to abstract the behavior of the mobile robot 210 with game control. Therefore, regardless of the form of the mobile robot platforms A similar action will be available.

3, the high-level controller network 320 receives, as an input state, an abstraction map 112, which has been preprocessed 310 as an input state, as an example of a network learned by deep reinforcement learning, And may provide information about the behavior for the robot 210. At this time, the controller abstraction (330) shows an example of a process of abstracting information about a behavior for the mobile robot 210 provided by the high-level controller network 320 into a game control signal. The motor controller 340 may control the motors included in the mobile robot 210 so that the mobile robot 210 executes an action corresponding to the abstracted game control signal.

The universal task of the mobile robot 210 is designed to evaluate the proposed methodology. In general, it is difficult to compare various modern mobile robots based on qualitative criteria. This is because each mobile robot has its own unique purpose and the hardware / software platform of the mobile robot is specialized according to its own purpose. However, there is a common goal of using mobile robots, even though the different purposes of different mobile robots are different, and the general performance of such mobile robots can be analyzed in a qualitative way.

For example, the performance of mobile robots can be broken down into functional modules such as cognition, world modeling, planning, performance tasks, and operations. These functions are the resultant functions of the robot operation regardless of the built-in arithmetic algorithm. A general evaluation framework for a mobile robot can be constructed by establishing a task to observe these functional modules simply and conveniently.

The mission designed for robotic performance evaluation is indoor multi-object tracking and gathering. The robot must search for, explore, and interact with the environment to perform the task. Specifically, the robot tracks and collects small objects that have a specific shape spread over the interior floor. The robot and the object are randomly placed in a room (for example, 3 m × 3 m in size). Two types of objects of different weights and sizes (for example, a 38 mm diameter ball and a 16 mm edge cube) are provided. These can be arranged to observe that the robot identifies the target and behaves differently.

4 is a diagram illustrating an example of a mobile robot according to an embodiment of the present invention. The robot shown in Fig. 4 is an indoor robot that tracks and collects a multi-object by itself. The robot is developed to demonstrate the functionality of the proposed system architecture and to show the general purpose of a mobile robot composed of recognition, search, and interaction . The chassis of this mobile robot is equipped with four mecanum wheels with holonomic movement and a single wide view camera (Genius, F100) is mounted on the platform for visual collection.

To ensure visibility, the platform and chassis are mechanically isolated with a silicon damper to ensure a very stable video stream. This structure can absorb camera shake or damage from camera and processor from physical impact due to rapid acceleration. In the case of end-effectors, a suction motor was used to show the interaction between the target and the robot.

Specifically, a 200W BLDC suction motor can be used to collect targets within 4cm from the nozzle tip. The inner radius of the nozzle can be controlled according to the size and weight of the detected target. 5 is a view showing an example of controlling the inner radius of the nozzle in an embodiment of the present invention. Compared to the left image, the right image shows an example in which the inner radius of the nozzle is reduced. The collection performance can be optimized through control of the inner radius of the nozzle. The mechanism behind the variable diameter nozzle is to increase the flow rate due to the reduction in diameter, so that an air bag for volume control can be formed between the coaxial cylindrical latex films along the inner and inner surfaces of the nozzle. To collect objects of various sizes and weights while maintaining the power of the suction motor constant, the robot can adjust the cross-sectional area of the nozzle to produce variable air flow pressures and velocities.

To implement the proposed software architecture, the robot consists of a small PC (Intel, NUC) and a mobile GPU processor (Nvidia, Jetson TX2). The main software runs in the ROS environment. Seven servomotors (Dynamixel MX-28AT) and suction motors (BLDC, 220W) are controlled by a real-time processor (NI, myRIO) using FPGA (Xilinx, Z-7010). PCs interface with LIDAR (Slamtec, RPLidar A2), IMU (EBIMU-9DOFV3) and Cameras for high-level control (Wide Cam (Genius 120 ° Wide)) and mobile GPUs handle deep neural networks for high- . All processors communicate via wireless communication. The specifications of the developed robot are described in Table 1.

Mechanical Specification Electrical Specifications Size [mm] 350 × 420 × 390 Battery [Wh] 240 (6S4P Li-ion) Weight [kg] 9.6 (7.9 *) Loading Min. Max. Speed [m / s] 0.76 Operating Time [Hr] 2.5 One Inhalation Range [mm] > 50 Power Load [W] 90 220

* Weight excluding battery

As previously mentioned, the software architecture is based on three main components: fast visual recognition 111, framework for asynchronous deep classification network 120 and GRL-planner (motion planner 113) and data between these major components And an abstraction map 112 serving as a data hub that integrates flows.

Estimates the absolute coordinates and size of the target object candidate in the raw camera image using an algorithm for fast visual recognition 111 as a component of the essential control loop 110. Robot implementations use OpenCV based color filtering and blob detection algorithm to recognize the position and size of each object in the camera image. The absolute coordinates of each object are estimated using a pre-computed transition matrix between the camera plane and the bottom plane at the recognized position of the camera image.

6 is a diagram illustrating examples of an object recognition screen and an abstract map screen according to an exemplary embodiment of the present invention. 6A shows an example in which the target objects detected in the camera view are marked as green boxes. FIG. 6B shows an example in which the positions of the target object and the obstacle are mapped to the abstract map 112 Respectively. The asynchronous deep classification network 120 may classify each target object, update the classification information in the abstraction map 112, and remove the false object non-dynamically from the abstraction map 112.

During the process of fast visual recognition 111, the cropped image of the target object candidate represented by the green box in Figure 6 (a) may also be stored as object templates for detailed recognition in the asynchronous deep classification network 120 .

The abstraction map 112 may combine the target object candidate location and size data of the fast visual recognition 111 with the obstacle location data of the LIDAR. For implementation, the abstraction map is designed to illustrate the typical size of a room with a resolution of 0.05m, 5m by 5m space. As shown in FIG. 6 (b), the object candidate and obstacle are mapped to the abstraction map 112, and for detailed object recognition using the asynchronous deep classification network 120, Can be tracked by the implemented multi-body tracker.

FIG. 7 is a diagram illustrating an example of a state diagram of a multi-body tracker including four states according to an embodiment of the present invention. The four states may be a detected, tracked, lost, or inactive state. When a target object is detected in the camera image, the detected object can be classified as a detected state.

(a) To classify the detected objects into registered targets and new targets, the multibody tracker can compare the detected targets with the currently registered targets. After predicting the location of the currently registered target with the estimated camera running history, the multi-body tracker can match each object based on the overlap of the boundary boxes of the camera screen. The unmatched target object is registered as a new tracked state object and the matching object can be used to update the template of the tracked object.

(b) If the tracked target objects do not match, the target can move to the lost state.

(c) If the object remains in a lost state for a period of time, it can move to the inactive state.

(d) When operating asynchronously, the asynchronous deep classification network 120 may continuously classify the registered target objects. If an object is classified as a false target, it can move to the inactive state regardless of its current state.

The asynchronous deep classification network 120 may classify the object object candidates in the abstraction map 112. In addition, the asynchronous deep classification network 120 can remove false objects that are mistaken in the course of fast visual recognition 111. [ For the asynchronous deep classification network 120, VGGNet-16, which has the most stable performance among the existing networks, was used. In VGGNet-16, the fully connected layer was replaced by two completely connected layers with 1024 units and the network was fine-tuned for this example. Among the software components, the asynchronous deep classification network 120 operates on a mobile GPU, Jetson TX2. The overall learning process was performed in a Keras environment with a TensorFlow backend.

In addition, a mobile robot control system similar to a video game environment is abstracted to implement a GRL-planner. Instead of using the raw image of the camera as an input state, an abstraction map 112 is created by using the positional data of the target object and the surrounding obstacles and used as the input state of the GRL-planar. Ten behaviors (eight directions in the body, left and right) to control the developed holonomic mobile robot were used as the outputs of the reinforcement learning network. The asynchronous advantage-critic (A3C) algorithm was used for reinforcement learning to show the state of art performance in video games. The network model is implemented in the same way as the A3C-LSTM model of DeepMind. The model uses four consecutive grayscale images as input, consisting of a single fully connected layer and two convolution layers followed by a 256-LSTM layer. The first convolution layer has a 32 8 × 8 filter with a stride of 4 and a second 64 4 × 4 filter with a stride of 2. There are 256 hidden units in a fully connected layer.

8 is a diagram showing an example of a simulation environment in an embodiment of the present invention. The simulation environment according to the present embodiment uses 10 joystick command input formations as in the holonomics platform of FIG. 3, and the center (robot position) of the red box 810 to guide the robot to collect the ball in front, The closer you touch the target, the higher the compensation is designed. Simulations can be terminated if the robot can not collect all targets or find targets at a specific time step.

The reinforcement learning network was learned in this simulation environment of Fig. The abstraction map 112 is downsampled at a resolution of 0.15 m to reduce the computational complexity. For a realistic simulation environment, object detection failure and object location inaccuracy are modeled as probabilistic flickering and shaking simulators. Overall learning was done using the TensorFlow framework, which took 6 hours to converge in a 6 CPU core environment.

Conventional image recognition algorithms are faster than deep learning based recognition algorithms, but the update rate is still slower than the control layer update rate. Thus, the IMU sensor was used to evaluate the robot's running measurements and to estimate the abstraction map 112 between target object updates. This driving record update has been able to maximize the bandwidth of the essential control loop 110.

Considering the size of the mobile robot (35cm × 39cm), a 45cm × 45cm area around the robot is designated as the task execution area. When the object comes in this area, the robot is operated to collect the object and the corresponding nozzle is arranged in the corresponding direction Move slowly toward the object. In addition, the cross-section of the variable shape nozzle is adjusted according to the object type recognized by the asynchronous deep classification network 120.

To summarize the performance of the designed robot, the asynchronous deep classification network 120 and the GRL-planner are robustly embedded and integrated into the robot to improve the functionality of the robot with stand-alone functionality. By utilizing the framework for the asynchronous deep classification network 120, the robot successfully interacted with various types of objects, as shown in FIG. In addition, the robot correctly searches the object using the learned GRL-planner in a simple simulation environment as described above.

9 is a diagram illustrating examples of the results of a robot path plan in an embodiment of the present invention. The green circle represents the target object, the blue arrow represents the expected path of the nearest first algorithm, and the yellow line represents the actual path of the robot acquired by the GRL-planner. This result shows that the performance according to the embodiments of the present invention surpasses the performance of the conventional heuristic algorithm for tracking the closest object.

In particular, the proposed architecture operates the entire system with a bandwidth of 50 Hz, and the overall operating bandwidth is not limited by the deep classification processing time. In order to analyze the proposed architecture and robots based on quantitative indices, the processing time and power consumption of each software component of the robot are measured as shown in Table 2.

Component Processor Computation Time [ms] Power Load [W] Recognition Intel i5-5600U 40 8.9 Obstacle Detection Intel i5-5600U 80 2.2 Deep Classifier TX2 Max-N (batch 4) 315 12.8 Motion Planner Intel i5-5600U 20 2.2 Odometry Estimation Intel i5-5600U 20 -

The robot was equipped with a high-performance GPU such as the Nvidia Titan X, and the table was evaluated assuming that it was in stand-alone mode of operation without a server. The update speed of the hardware control layer depends only on the computation time of the motion planner 113 by separating the essential control loop 111 from other software components such as the fast visual recognition 111 or the deep classifier 240. [ The asynchronous deep classification network 120 operates on the mobile GPU Jetson TX2 with a maximum GPU frequency mode and a batch size of 4 to classify four objects in one cycle. However, if robots can access the additional computational capabilities of a local server with a high-performance GPU, the performance of the robot can be further improved. The improved results for the processing time of the deep classifier 240 are shown in Table 3.

Processor Computation Time [ms] batch 1 batch 4 batch 8 batch 16 Intel i5-5600U 1560 5800 11000 - Jetson TX2 Max-P 125 360 710 1430 Jetson TX2 Max-N 110 315 590 1230 Nvidia GTX 960 30 (40 *) 92 (100 *) 175 (195 *) 340 (365 *) Nvidia Titan X 25 (35 *) 82 (91 *) 163 (183 *) 295 (380 *)

* Computation time including communication delay

However, even if an external GPU server is accessed, the processing time of the asynchronous deep classification network 120 is not fast enough to operate the robot in real time if the asynchronous deep classification network 120 is included in the essential control loop 110 of the robot. In other words, asynchronous deep classification network 120 is excluded from the essential control loop 110 of the mobile robot and processed in an asynchronous manner, resulting in a processing time fast enough to operate the mobile robot in real time.

10 is a block diagram showing an example of the internal configuration of a mobile robot according to an embodiment of the present invention. 10, the mobile robot 1000 according to the present embodiment includes a memory 1010, a processor 1020, an input / output interface 1030, a camera 1040, a sensor 1050, and a motor 1060 . The memory 1010, the processor 1020 and the input / output interface 1030 may be implemented as separate computer devices for controlling the mobile robot 1000 included in the mobile robot 1000. A plurality of processors 1020 may be used.

The memory 1010 may be a computer-readable recording medium and may include a permanent mass storage device such as a random access memory (RAM), a read only memory (ROM), and a disk drive. Here, the non-decaying mass storage device such as the ROM and the disk drive may be included in the mobile robot 1000 as a separate persistent storage device different from the memory 1010. The memory 1010 may also store an operating system and at least one program code. These software components may be loaded into the memory 1010 from a computer readable recording medium separate from the memory 1010. [ Such a computer-readable recording medium may include a computer-readable recording medium such as a floppy drive, a disk, a tape, a DVD / CD-ROM drive, and a memory card. In other embodiments, the software components may be loaded into memory 1010 via a separate communication interface rather than a computer-readable recording medium. For example, the software components may be loaded into the memory 1010 of the mobile robot 1000 based on a computer program installed by files received via the network. In this case, the mobile robot 1000 may further include a separate communication interface for communicating with other computer devices via an external network.

The processor 1020 may be configured to process instructions of a computer program by performing basic arithmetic, logic, and input / output operations. The instructions may be provided to the processor 1020 by the memory 1010. For example, processor 1020 may be configured to execute instructions received in accordance with program code stored in a recording device, such as memory 1010. [

The input / output interface 1030 may be a means for interfacing with other components such as the camera 1040, the sensor 1050 and the motor 1060. For example, it has been described that the components described with reference to FIG. 5 may be connected to each other through wireless communication. Specific embodiments for the camera 1040, the sensor 1050 or the motor 1060 and the like have already been described, and thus repetitive description will be omitted.

Further, in other embodiments, the mobile robot 1000 may include fewer or more components than the components of FIG. However, there is no need to clearly illustrate most prior art components. For example, the mobile robot 1000 may be implemented in accordance with the operation purpose of the mobile robot 1000, and may be configured to perform a specific operation (for example, rotation of a wheel, intake of air using the above- (For example, a nozzle for sucking an object in the above-described embodiment) for performing the above-described operation.

11 is a flowchart showing an example of a target classification method according to an embodiment of the present invention. The target classification method according to the present embodiment can be performed according to the control of the processor 1020 of the mobile robot 1000 described above.

In step 1110, the mobile robot 1000 performs image processing on the image input through the camera 1040 included in the mobile robot 1000, and outputs the image of the target region corresponding to the target in the image and the image of the region of interest Information can be obtained.

In step 1120, the mobile robot 1000 may map the position of the target on a two-dimensional plane to generate an abstraction map. The generation of such an abstraction map has been described in detail with reference to FIG.

In step 1130, the mobile robot 1000 may store image information for a region of interest corresponding to a target in the image, corresponding to the location of the mapped target in the abstraction map. 6, an example of storing image information of a region of interest corresponding to a position of a target has been described.

In step 1140, the mobile robot 1000 may process the image information of the ROI stored in the abstraction map through the asynchronous Deep Classification Network to classify the target asynchronously. In this case, the mobile robot 1000 may perform movement of the mobile robot 1000 or the target by multi-body tracking (hereinafter, referred to as " multi-body tracking ") in order to eliminate mismatch of information about the target stored in the abstraction map according to the movement of the mobile robot 1000 or the target. multi-body tracking algorithm.

In step 1150, the mobile robot 1000 may update the abstraction map with information about asynchronously sorted targets via the asynchronous deep classification network. The abstraction map can be used as an input state of the motion planner, and the motion planner can determine the motion of the mobile robot 1000. At this time, as information about the asynchronously classified target is updated in the abstraction map, information about the asynchronously classified target can be reflected in the determination of the motion 1000 of the mobile robot.

Also, the mobile robot 1000 can abstract the motion of the mobile robot 1000 determined as the output of the motion planner using at least one of a predetermined plurality of game control signals, The motion of the mobile robot 1000 can be controlled by controlling the motor 1060 included in the mobile robot 1000. [ The control of the mobile robot 1000 will be described in more detail with reference to FIG.

12 is a flowchart illustrating an example of a mobile robot control method according to an embodiment of the present invention. The mobile robot control method according to the present embodiment can be performed under the control of the processor 1020 of the mobile robot 1000 described above. Steps 1240 to 1280 of FIG. 12 can also be applied to the embodiment of FIG. In addition, the process of updating the abstraction map in Fig. 11 may be applied to the embodiment of Fig.

In step 1210, the mobile robot 1000 may acquire information about the target through image processing on the image input through the camera 1040 included in the mobile robot 1000. [

In operation 1220, the mobile robot 1000 may acquire sensing data output through at least one sensor 1050 included in the mobile robot 1000. The sensing data may include, for example, an output of a rider or an IMU.

In operation 1230, the mobile robot 1000 may generate an abstraction map by mapping information about the target and sensing data on a two-dimensional plane. The generation of such an abstraction map has been described in detail with reference to FIG. Step 1230 describes an embodiment in which an abstraction map is generated by further using sensing data as compared to step 1120 in Fig.

In step 1240, the mobile robot 1000 may determine the motion of the mobile robot 1000 by using the generated abstraction map as an input state of the motion planner.

In step 1250, the mobile robot 1000 may construct a simulation environment based on the generated abstraction map.

In step 1260, the mobile robot 1000 can reinforce the motion planner through the established simulation environment. The simulation environment may correspond to the simulation environment of the two dimensional video game and the mobile robot 1000 may reinforce learning of the motion planner using an algorithm for gaming reinforcement learning (GRL) in step 1260 .

In step 1270, the mobile robot 1000 may abstract the motion 1000 of the determined mobile robot using at least one of a predetermined plurality of game control signals.

In step 1280, the mobile robot 1000 may control the motion of the mobile robot by controlling the motor 1060 included in the mobile robot 1000 according to at least one signal.

As described above, according to the embodiments of the present invention, unlike the end-to-end robot software architecture, an asynchronous deep classifying network capable of introducing a deep learning technology as one component constituting the entire software of the mobile robot is used It is possible to effectively integrate large-scale deep neural networks into mobile robots that require bandwidth control. In addition, it can reduce learning complexity and improve algorithm scalability by using gaming enhanced learning based motion planner for mobile robots.

The system or apparatus described above may be implemented as a hardware component, or a combination of hardware components and software components. For example, the apparatus and components described in the embodiments may be implemented within a computer system, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA) , A programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications running on the operating system. The processing device may also access, store, manipulate, process, and generate data in response to execution of the software. For ease of understanding, the processing apparatus may be described as being used singly, but those skilled in the art will recognize that the processing apparatus may have a plurality of processing elements and / As shown in FIG. For example, the processing unit may comprise a plurality of processors or one processor and one controller. Other processing configurations are also possible, such as a parallel processor.

The software may include a computer program, code, instructions, or a combination of one or more of the foregoing, and may be configured to configure the processing device to operate as desired or to process it collectively or collectively Device can be commanded. The software and / or data may be in the form of any type of machine, component, physical device, virtual equipment, computer storage media, or device As shown in FIG. The software may be distributed over a networked computer system and stored or executed in a distributed manner. The software and data may be stored on one or more computer readable recording media.

The method according to an embodiment may be implemented in the form of a program command that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, and the like, alone or in combination. The medium may be one that continues to store computer executable programs, or temporarily store them for execution or download. In addition, the medium may be a variety of recording means or storage means in the form of a combination of a single hardware or a plurality of hardware, but is not limited to a medium directly connected to a computer system, but may be dispersed on a network. Examples of the medium include a magnetic medium such as a hard disk, a floppy disk and a magnetic tape, an optical recording medium such as CD-ROM and DVD, a magneto-optical medium such as a floptical disk, And program instructions including ROM, RAM, flash memory, and the like. As another example of the medium, a recording medium or a storage medium managed by a site or a server that supplies or distributes an application store or various other software to distribute the application may be mentioned. Examples of program instructions include machine language code such as those produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter or the like.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments. For example, it is to be understood that the techniques described may be performed in a different order than the described methods, and / or that components of the described systems, structures, devices, circuits, Lt; / RTI > or equivalents, even if it is replaced or replaced.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims (15)

Obtaining image information about a position of a target and a region of interest corresponding to the target in the image through image processing of an image inputted through a camera included in the mobile robot;
Mapping the position of the target on a two-dimensional plane to generate an abstraction map;
Storing image information for a region of interest corresponding to the target in the image, corresponding to a position of the target mapped to the abstraction map;
Processing image information for a region of interest stored in the abstraction map through an asynchronous Deep Classification Network to classify the target asynchronously; And
Updating information on the asynchronously classified target through the asynchronous deep classification network to the abstraction map
Wherein the target classification method comprises: The method according to claim 1,
Determining the motion of the mobile robot using the abstraction map as an input state of the motion planner
Further comprising:
And information about the asynchronously classified target is reflected in the determination of the motion of the mobile robot as information about the asynchronously classified target is updated in the abstraction map. 3. The method of claim 2,
Abstracting the motion of the mobile robot determined as the output of the motion planner using at least one of a predetermined plurality of game control signals; And
Controlling a motion of the mobile robot by controlling a motor included in the mobile robot according to the at least one signal
Further comprising the steps of: 3. The method of claim 2,
Constructing a simulation environment based on the generated abstraction map; And
The step of reinforcement learning of the motion planner through the constructed simulation environment
Further comprising the steps of: 5. The method of claim 4,
The simulation environment corresponds to a simulation environment of a two-dimensional video game,
The step of reinforcement learning of the motion planner includes:
Wherein the motion planner is reinforcement learning using an algorithm for gaming reinforcement learning (GRL). The method according to claim 1,
Wherein the asynchronously classifying comprises:
Wherein the movement of the mobile robot or the target is detected using a multi-body tracking algorithm in order to eliminate mismatch of information about the target stored in the abstraction map according to movement of the mobile robot or the target Wherein the target is classified into two categories. A computer program stored in a computer-readable medium for causing a computer to execute the method of any one of claims 1 to 6 in combination with the computer. A computer-readable recording medium storing a program for causing a computer to execute the method according to any one of claims 1 to 6. A computer apparatus for controlling a mobile robot,
A memory for storing instructions readable by a computer; And
At least one processor < RTI ID = 0.0 >
Lt; / RTI >
Wherein the at least one processor comprises:
Acquiring image information about a position of a target and a region of interest corresponding to the target in the image through image processing of an image input through a camera included in the mobile robot,
Mapping the position of the target on a two-dimensional plane to generate an abstraction map,
Storing image information for a region of interest corresponding to the target in the image corresponding to a position of the target mapped to the abstraction map,
Processing the image information for a region of interest stored in the abstraction map through an asynchronous Deep Classification Network to classify the target asynchronously,
Updating the abstraction map with information about asynchronously classified targets via the asynchronous deep classification network
The computer device comprising: 10. The method of claim 9,
Wherein the at least one processor comprises:
Determining the motion of the mobile robot using the abstraction map as an input state of the motion planner,
Wherein information about the asynchronously classified target is reflected in the determination of the motion of the mobile robot as information about the asynchronously classified target is updated in the abstraction map. 11. The method of claim 10,
Wherein the at least one processor comprises:
Abstracting the motion of the mobile robot determined as the output of the motion planner using at least one of a predetermined plurality of game control signals,
And controls the motion of the mobile robot by controlling a motor included in the mobile robot according to the at least one signal. 11. The method of claim 10,
Wherein the at least one processor comprises:
Constructing a simulation environment based on the generated abstraction map,
Reinforcing the motion planner through the constructed simulation environment
The computer device comprising: 13. The method of claim 12,
The simulation environment corresponds to a simulation environment of a two-dimensional video game,
Wherein the at least one processor comprises:
Wherein the motion planner is reinforcement learning using an algorithm for gaming reinforcement learning (GRL). 10. The method of claim 9,
Wherein the at least one processor comprises:
Wherein the movement of the mobile robot or the target is detected using a multi-body tracking algorithm in order to eliminate mismatch of information about the target stored in the abstraction map according to movement of the mobile robot or the target Said computer system comprising: In a mobile robot,
A memory for storing instructions readable by a computer;
At least one processor configured to execute the instruction; And
Camera receiving images
Lt; / RTI >
Wherein the at least one processor comprises:
Acquiring image information about a position of a target and a region of interest corresponding to the target in the image through image processing on an image input through the camera,
Mapping the position of the target on a two-dimensional plane to generate an abstraction map,
Storing image information for a region of interest corresponding to the target in the image corresponding to a position of the target mapped to the abstraction map,
Processing the image information for a region of interest stored in the abstraction map through an asynchronous Deep Classification Network to classify the target asynchronously,
Updating the abstraction map with information about asynchronously classified targets via the asynchronous deep classification network
And a mobile robot.

Images