DE4001493A1 - Automatic control of vehicle motion - using ultrasonic object sensor providing input to learning process for vehicle control - Google Patents

Abstract

An automatic motion control system, for a vehicle or mobile robot, uses signals generated by ultrasonic sensors (4) that are located on the surface to detect the presence of any object. The sensor signals are fed to a preprocessor (10) for binary coding that generates (27) binary outputs. The signals are combined in a neutral network and are received by a converter (20) coupled to the vehicle drives. The vehicle is moved in a selected environment as part of a learning process. Data is entered into memory for later use. ADVANTAGE - Improved motion control and collision avoidance.

Description

The invention relates to a method and a device for automatic control or control support of autonomously or partially autonomously movable devices such as driving witnesses, robots etc., with the characteristics of the generic term of claim 1.

In a known method for processing the Sensor signals used a conventional computer, the on the principle of "von Neuman-based", which is based on Calculator, tail and memory is provided and over Programs is controlled.

For complex real-time tasks such as sensor data processing device for autonomous control of a vehicle or Robotic arm has the use of such a conventional one Calculator the following disadvantages:

• 1. The data must be in serial information processing controlled by a computer program by the arithmetic unit managed and offset. This is time consuming.
• 2. Program creation is complicated and often very difficult time consuming.
• 3. The system knowledge must be explicitly known beforehand.
• 4. Only linear relationships can be mapped. Synergistic or complex cybernetic systems cannot be mapped.  
• 5. The controller is sensitive to sensor malfunctions data, the processing program and against the device fall (no fault tolerance).

An information processing network structure, a so-called called "neural network", denotes an information processing processing system, which is made of parallel working processo ren exists, which in principle arbitrarily with each other, mostly but are linked from a hierarchical point of view, the processors being both simple and complex Calculations can perform the linking of the processes sensors among themselves via multiplicative connections follows which over the duration of information processing adjustable, that is to say changeable, and certain Processors of the system take on the designated task, exchange information with the environment.

So far, neural networks have been the main subject theoretical investigations with the aim of due to the "von Neuman" architecture of today's computers set boundaries to be broken (essay "Natural and Art Intelligence "by R. Opitz DE-Z Forum (Data General GmbH, 1984, pp. 15-18)). With the help of structural calculators adaptive, neural networks, which are numerous have connected nodes, so-called neurons, Researchers are currently looking for functions and properties worldwide human brain, to simulate and ultimately replicate it. Pragmatic implementations However, due to a lack of understanding and of trans difficulties of the cybernetic connections to be found. So there are no practical applications in industrially manufactured products known.

The invention has for its object a method and a device specified in the preamble of claim 1  to create the same type in which the automatic control or control aid of devices that are in a minimum an obstacle-bearing environment and one Have a driving or movement task with the aim of one collision-free avoidance of the obstacle or obstacle nisse and a shunting tasks adapted to the environment lung can be improved.

A method using the Features of claim 1 and a device with the Features of claim 13.

Advantageous embodiments of the invention are in the Subclaims specified.

The invention is particularly advantageous for automatic operation Control of vehicles as well as driving or shunting help usable. The vehicle control with a procedure and a device according to the invention enables namely

• - avoiding an obstacle collision when driving autonomously, where a parallel trip to objects, a passage have gates and the like realized;
• - a driving aid for the vehicle normally leading drivers, for example when parking, at the Short-range navigation or when reversing;
• - Improved situation detection through classification an obstacle or object contour;
• - Cooperative behavior during overtaking and evasive maneuvers vern;
• - Avoiding rear-end collisions.

The application of a method and a facility according to the invention not only enables advantageous solutions conditions in road traffic but also of transport tasks in Factories, at train stations and airports as well as in construction sites traffic with complex unmanageable environmental conditions.

Due to the at least temporarily autonomous vehicle control human misbehavior when driving largely turn off.

However, the invention can also be used with other movable devices, how robots are used where the control of the Robots move around using conventional computers is time consuming and time consuming and always one programming, especially in hostiler, chaotic and dynamically changing or from the front in unknown surroundings.

Programming is not necessary when using a neural Network. It is possible for the purposes of the invention a conventional von Neuman computer to simulate a neural network. To provide higher However, real-time capabilities are advantageous instead of one conventional computers used "real" neural computers, which have so-called neurochips.

The invention is in the following with reference to drawings further details explained.

Show it:

Figure 1 is a schematic of an equipped with a controller according to the invention autonomous or semi-autonomous vehicle.

Fig. 2 is a diagram of simulation of the learning phase of an autonomous vehicle guidance;

Fig. 3 is a schematic of the simulation can phase of an autonomous vehicle guidance;

FIGS. 4 and 5 screens during Configurati phase and the learning phase;

FIGS. 6 and 7 screens before the start of the learning phase, and in training the neural network;

Fig. 8 is a combined screen display during the can phase;

Fig. 9 shows an example of the structure of a neural network according to the invention for autonomous vehicle management in traffic.

In the scheme of Fig. 1 nine ultrasonic transducers 4 are arranged with Meßkeulen 8 on a vehicle 2 in total, whose output signals are supplied to the inputs 8 of a preprocessor 10 for binary coding. At the outputs 12 of this preprocessor 10 there are twenty-seven output signals in binary form. These twenty-seven output signals are input to an input interface of a neural network designated as a whole by reference numeral 14 and from there via the neurons denoted by 16 in the form of output signals 18 to a converter device 20 which converts the output signals into control signals 22 , the number thereof corresponds to the number of driving functions to be performed (acceleration / braking). These signals are entered into actuators (brakes / drives) of the vehicle designated as a whole with the reference numeral 24 for triggering the desired driving function, as indicated in FIG. 1 with a fan of arrows.

The driving function triggered by this leads to new sensors  signals that are reprocessed as described.

FIG. 2 shows an overview diagram of a simulation of the control which can be carried out with a system according to FIG. 1 by means of two personal computers PC 1 and PC 2 .

In the block designated by 26 in FIG. 2, a so-called driving reason is created on the screen of the PC 1 by building up obstacles in a two-dimensional field.

In block 28 , a learning file is created by externally controlled guidance of the vehicle through the driving ground along a selected path to be learned.

In block 30 , the neural network 14 according to FIG. 1 is trained with the learning file on the screen of the PC 2 .

In Fig. 3, the driving of the vehicle is shown in the "can phase" schematically. Here the two personal computers PC 1 and PC 2 are interconnected very interactively. On the screen of the PC 1 , as shown in block 32, it is determined how the vehicle is traveling autonomously, ie controlled by the neural network 14 , in the driving reason set with block 26 . In block 34 , the trained network and the sensors on the screen of the PC 2 are Darge.

In Figs. 4 and 5, the screen images according to blocks 26 and 28 in Fig. 2 in detail.

In Figs. 4 and 5 obstacles by dark-colored circles 27, rectangles 29, 31 and an angle 33 are symbolized. These obstacles are placed according to FIG. 4 to the configuration shown, ie a "driving reason" by operating a mouse, not shown. Keys 38 to 46 are drawn under the actual screen 36 of the PC 1 and can be tapped to perform the "Load", "Save", "New", "Help" and "Back" functions.

Existing driving reasons can be changed. That’s it Call up the "Load" field and display the name of the driving reason give.

With "Save" modified or newly created Driving reasons are saved.

With the button "new" the area can be deleted at any time and a new driving reason can be created.

By pressing the "back" button, the configurati ons phase and jump to the main menu.

According to the illustration in FIG. 5, after creating a driving reason on the screen 36 of the PC 1, a simulated vehicle 50 is placed in the driving reason and controlled externally along a learning course through the driving reason, which is indicated by means of a dash-dotted line 52 in FIG. 5 .

In this case, the functions "Load", "Start", "Stop", "Help" and "Back" are set using keys 38 to 46 .

The direction of travel into which the vehicle 50 is externally controlled on the screen by means of a wind rose can be read in the field 37 shown on the right of the screen 36 .

Before a neural network can be trained, it must be dimensioned and parameterized. Since nine sensors 4 are provided on vehicle 2 and their values are coded with three bits, the size of the input vector and thus the number of neurons 16 at the input interface is a total of twenty-seven. In the example shown, the output vector consists of four bits, although eight driving possibilities are coded in the unit 20 .

It is therefore possible to include the topology so-called "hidden levels" to vary, as well as the Change learning parameters. These parameters are in the neuro nal simulator entered in a table. In this In the table, a coefficient I corresponds to the learning parameter η and a coefficient II the moment term α. For an opti Male choice of parameters η, α are used in the literature for other cases the values η = 0.9 and α = 0.6 are recommended. Other Experience has shown that values can convert to faster lead.

The implementation of a learning phase on the neural simulator is explained below with reference to FIGS. 6 and 7. Initially, a neural network with twenty-seven input neurons, four output neurons and any number and size of hidden levels is built up in the simulator. This network is initialized with random values from a small range of values (-0.1 / 01). The learning process can now be started. The input neurons are denoted in block 30 (according to FIG. 2), which symbolizes the screen of the PC 2 , with the reference number 56 and the output neurons with the reference number 58 . An OK key 60 and a delete key 62 are shown below the input neurons 56 . Three windows 64 , 66 and 68 are shown below, the window 64 , 66 indicating that a convergence file and an information file are being created. The name of the learning file appears in window 68 for confirmation. The learning process is started by clicking the OK button.

The convergence file is used to run the learning process in case of longer learning periods (2 hours to 3 days) to be able to pull. After each learning cycle, i.e. H. after complete dig through the learning file, the global Network errors, the number of cycles and the number of Fields entered during the last run of the learning file.

The convergence file can last up to 100 for longer learning periods KB grow big. Using the convergence file, an over pace local minima as well as the rate of convergence be recognized.

In any case, the information file consists of only four lines. These lines are entered in the information file again after each learning cycle. The content looks e.g. B. from:
Number of learning steps - 60 792
Number of cycles - 447
Number of samples - 136
global error - 0.057

This file gives the current learning status of the neuro network. This allows you to cancel the Learning process and a later training of the network.

In Fig. 7 the training of the neural network is Darge presents. If the OK button 60 was pressed, the display on the screen 30 is approximately as follows:

In the part below the input neurons 56 , the number of patterns, the number of representations, the number of cycles (passes through the learning file), the number of errors in the current cycle and the global error, which is recalculated after each cycle, shown in line 70 . Values read from the learning file and encoded in a binary input vector are displayed in sensor line 72 . Above the network, the current output of the network is displayed in field 74 and the desired, ie the output to be trained, is displayed in field 76 .

In general, the global error will steadily decrease. When the global error is very close to zero and doesn't change over several cycles, the learning process can be canceled. The Errors per cycle are not necessarily zero. With very large ones Learning files (more than 500 patterns) can contradict themselves learning patterns. In this case, the global Errors converge to a number between three and ten and still have some errors per cycle. these can selected and corrected. Nevertheless, the network is off adequately trained and can be tried in the can phase will. The learning process is not shown using a ten "Escape" key canceled.

In Fig. 8, the optional phase is shown by interactive interaction between the two personal computers PC 1 and PC 2 and the representation of their two screens 28 and 30 .

By pressing the key 38 , a desired driving reason is displayed on the screen of the PC 1 and the vehicle 50 is positioned as shown.

On the PC 2 , the simulator is started and the neural network trained for driving in accordance with PC 1 is loaded and started, resulting in the screen structure shown on the right in FIG. 8. The simulation is started by pressing key 40 .

On the screen 36 of the PC 1 , the vehicle 50 can be seen to drive autonomously through the neural network on the PC 2 around the obstacles. If the obstacles are cut or run over, this means that

• - the network has not been adequately trained,
• - the learning file contained too few learning patterns,
• - the learning file too many contradicting input-output Had relations,
• - the wrong driving reason has been loaded.

Because by repositioning the vehicle and starting again Another path can be found, some should Tried out starts at different starting positions will.

Even if perfect driving behavior is not expected can, but should in principle to the built obstacle nissen-oriented driving behavior, such as reflexive avoidance at the obstacles, become obvious.

Line 70 on the screen of PC 2 shows the current sensor values, the coded input vector, the output vector and the current direction of travel of vehicle 50 .

The optional phase is ended by pressing the stop button 42 .

In the example described, the autonomous driving of a vehicle 50 after learning to drive around an obstacle is shown on a predetermined driving surface ( FIG. 4).

However, the invention can also be applied to several such Vehicles for training cooperative driving behavior use multiple vehicles. It is after creating a driving reason and a learning file generation as well  a trained network for training on the learning vehicle Aim to control all vehicles, with the vehicles freely selectable and collision-free simultaneous autonomous driving of all vehicles the final Learning objective is.

An example of the structure of the overall architecture of a learning vehicle with a neural network is shown in Fig. 9 Darge. It includes driving strategy, route planning, object recognition, maps and a learning module.

Here, in the two lower oval fields 80 , 82 sen sensors for the overall position of the vehicle and distance measuring sensors as well as 84 actuators for actuating the various driving functions are indicated in the upper oval field.

Field 86 contains a map with environmental obstacles, while field 88 contains a map with other environmental properties.

An obstacle classifier is indicated in field 90 . Fields 86 , 88 , 90 are followed by fields 92 , 94 , 96 . Field 92 contains a module for learning according to an overarching learning goal. Field 94 contains an optimization device for route planning. Field 96 contains a module for developing driving strategies. A control device for associative actuator control is arranged between the actuator field 84 and the fields 94 , 96 .

With the described method and the described A direction according to the invention can be the following advantage achieve adhesive effects:

• 1. When training a simulation vehicle with sensors, Actuators and a neural network turned out  out that driven tracks in obstacle courses partially were abstracted (reflex behavior) and after recall a similar driving behavior by the Network was achievable for the vehicle.
• 2. It was not, as previously necessary, vehicle controls and driving behavior with considerable effort to pro gramming.
• 3. The fault tolerance against sensor faults was very high.
• 4. In addition to training individual obstacle rides were very complex driving situations without mathematical preparation learnable.
• 5. When using multiple vehicles with controls achievable through a network of cooperative driving behavior (Evasive maneuvers, driving strategies).
• 6. It could be a completely new type and quality of Power generation through "learning from examples" using a learning phase (training with examples) and an optional phase (retrieval of information from the Network) and information storage can be achieved.
• 7. The coverage of complex information representations is very high.
• 8. The polling time constant is in the millisecun range the. This is a factor higher than conventional AI Systems and is suitable for various real-time processes tasks of regulating and controlling cybernetic Processes.
• 9. Learning according to the example described can by  Learning according to a higher learning goal (reinforcement Learning) and an experimental module. In order to system architectures emerge.
• 10. The simulated networks can be trained in neuro VLSI implementations implemented and in products be used.

Claims (15)

1. A method for the automatic control of autonomously or partially autonomously movable devices with the environment on receiving location sensors for emitting sensor signals which are characteristic of the distance of the device from at least one obstacle or object, and with a processing unit into which the sensor signals are input are, and which at least emits a control signal for drives / brakes of the device for accelerating / braking the respective movement of the device in order to avoid a collision with the obstacle and to generate desired driving and movement behavior, characterized in that the processing unit as information processing network structure ("neural network") is formed or simulated that the device in a learning phase from a limited number of selected situations externally controlled executes a motion cycle that the sensor and actuator data records of the network structure obtained in egg ner learning phase repeatedly entered and the movement cycle are instructively specified, and that when falling below a predetermined global error, the learning phase is completed and a transition is made to a phase in which the device in a known or unknown environment autonomously influenced the or an unknown environment by the objects learned movement cycle at least approximates. 2. The method according to claim 1, characterized records that in a data acquisition phase the sensor / actuator data records with each execution of the externally controlled movement sequence cycle into one Learning file can be saved. 3. The method according to claim 2, characterized records that in each learning cycle during the Learn phase the learning file completely several times and learn go through properly compiled (to avoid of unlearning) and the number of cycles, the global Errors and the number of errors when running the Learning file are entered in a convergence file, and that when a predetermined conver is exceeded limit the learning cycle is ended. 4. The method according to claim 2 or 3, characterized ge indicates that a Information file containing the number of learners steps, the number of cycles, the number of learning pattern (driving reasons) and the global error becomes. 5. The method according to any one of the preceding claims, characterized in that the device a vehicle and the motion cycle a driving cycle along a path with the vehicle in one at least one environment with obstacles (driving reason) are. 6. The method according to any one of claims 1 to 4, characterized characterized that the device is a Robo ter and the motion sequence cycle is kinematic three-dimensional process along a trajectory  the robot in at least one obstacle send environment (driving reason). 7. The method according to any one of the preceding claims, characterized in that the Or tion sensors are each problem isomorphic and the sensor signals obtained from all sensors for the information processing network structure together to at least one control signal (input vector) be prepared. 8. The method according to any one of the preceding claims, characterized in that both Objects or object contours as well as object classifications network training together as Input size (vector) and the Generate an image of the actuator data records. 9. The method according to any one of the preceding claims, characterized in that further self-paced learning by default goal (learning sentences) through insertion and use a trial and exercise cycle with limited Number of attempts is made and positive Responses are saved on the network while negative behavior expires (self-adaptation). 10. The method according to any one of the preceding claims, characterized in that global Movement strategies and processes with local / situational trained behavioral representations combined and shares in the input vector of the network the movement strategy to be trained with that in the optional phase through higher hierarchies in the sense of hierarchical network architectures set up and  be set. 11. The method according to any one of the preceding claims, characterized in that Movements using past time gradual previous movement patterns with weight levels in the training phase in the input vector be used and procedures in the can phase the movement are taken into account and generated. 12. The method according to any one of the preceding claims, characterized in that convention computer with simulated neural networks, Hybrid systems with specially adapted elements and / or neurochips can be used. 13. Self-supporting or supportive facility Control of autonomously or partially autonomously movable Devices, such as vehicles, with the environment Location sensors for forming sensor signals, which for the distance of the device from at least one Hin dernis, an object contour or the like. characteristic and with a processing unit into which the sensor signals are entered, and what tax signals for drives / brakes of the device for acceleration nig / slow down the respective movement of the device issues, characterized in that the Processing unit an information processing Network structure (neural network) with one of the number number of input signals corresponding to the sensor signals Interfaces, a variety of neurons (nodes score) and one of the number of drive / brake radio corresponding number of output interfaces for the at least one control signal.   14. The device of claim 13, wherein the sensors on are attached to a vehicle that is autonomous in should move a given environment, thereby characterized in that the sensors as both sending and receiving sensors, esp especially as ultrasonic sensors or microwaves sensors that are designed in sync with a Cycle times are operated, which are the same or greater than the given greatest obstacle distance Corresponding term (sum of the outward and return runs time) of a sensor pulse at a given maximum vehicle speed. 15. The device according to claim 13 or 14, characterized ge indicates that the number of neurons at the input interfaces equal to the product the number of sensors multiplied by the information number (e.g. 3 bits) in parallel per sensor signal.