Autonomous mobile robot

from Wikipedia, the free encyclopedia

As autonomous mobile robots are robots referred to, which can move independently in their environment and operate. There are currently different levels of autonomy , i.e. the independence of the robot. Mobile robots are often referred to as autonomous if the software / electronics / hardware that controls them is located on the robot. The robot is then autonomous as long as its energy supply allows it. Giving the robot instructions on how or which task to perform does not disrupt its autonomy. A robot is only fully autonomous when the robot is also independent in terms of its energy supply, e.g. B. by means of a supply of energy for charging the batteries via solar cells. The development of mobile autonomous robots is being promoted through competitions such as the DARPA Grand Challenge or the RoboCup . General Motors is planning to test the first unmanned cars from 2015 and in series production from 2018.

definition

There is no standard definition for autonomous mobile robots, but the name already describes the most important criteria for such robots.

robot

Robots are stationary or mobile machines that perform defined tasks according to a specific program .

mobility

In the field of robotics, mobility means that the robot is not permanently mounted at a certain point, but can move more or less freely in a certain environment. The type of terrain determines the necessary actuators for movement. For example, while wheels are more suitable for flat terrain, chains or legs (see walking robots ) are usually advantageous for uneven terrain .

autonomy

The autonomy of a robot can be realized in different degrees. For example, robots that independently explore their surroundings and complete the tasks assigned to them, but have to return to a charging station again and again, are not completely autonomous. The necessary degree of autonomy that is necessary for a particular robot can be determined from the specification of the tasks to be performed.

Areas of application

Mobile robots differ greatly in design and structure due to their different areas of application:

Do-it-yourself and play robots

Autonomous small robots which are used in non-commercial hobby and play areas. Typical representatives are organized interest groups that also operate on an international scale.

Service robot

This commercial use of mobile robots is becoming increasingly important. Typical applications are: information platforms in hardware stores, inspection robots on oil platforms or in pipelines, service robots in care and domestic areas. Even if there is very intensive development in these areas, inserts in the market have not really caught on.

Industrial robots

Driverless transport vehicles are widely used in the industry. With the help of sensors and software for localization, navigation and path planning, mobile transport robots are available that can search for routes independently and adapt dynamically. If an industrial mobile robot is equipped with a robotic arm, it is called a mobile manipulator . Such systems are used, for example, in order picking .

hardware

The hardware of mobile autonomous robots can mainly be classified into two categories, sensors and actuators . In addition there is the control electronics , which can range from relatively simple circuits to complex computer systems . All of these components are mounted on a frame adapted to the special requirements of the respective robot.

Sensors

The sensor technology of autonomous mobile robots can be divided into two categories. Internal sensors supply the robot's control software with data about the robot's condition, while external sensors supply data about the environment. Internal sensor data would be, for example, data on the state of charge of the batteries, whereas data from external sensors include, for example, all data on the type and distance of obstacles.

Examples of internal sensors

  • Voltage measuring device (AD converter)
  • Odometry ("odometer")
  • Thermometer for overheating components
  • Radio receiver

Examples of external sensors

Actuators

Actuators are the counterpart to sensors and are used to manipulate the robot or the environment. Like sensors, actuators are divided into internal and external actuators. While the internal actuators change the state of the robot (for example by charging the batteries), external actuators are used, among other things, to move around or to move objects.

Examples of internal actuators

  • charger
  • fan
  • Heating (in very cold environments)
  • Relay for switching redundant modules
  • Channel

Examples of external actuators

  • Engines
  • Servos
  • pump
  • Lamps (for display purposes or as lighting for cameras)

software

There are various software approaches for realizing autonomous behavior. The following list of software architectures includes four important approaches for realizing autonomy in mobile robots. In practice, combinations of these techniques are often used and methods from the field of neural networks are included.

Behavior-based architecture (reactive architecture)

The reactive model is the simplest software architecture for autonomous robots (and software agents too ). The behavior of the robot is defined by a collection of stimulus-response (sense-act) rules. The sensory inputs are iterated over each of these rules. The outputs of these rules ( behaviors ) then determine the reaction of the robot, whereby a combination of the triggered behaviors is carried out. It is essential for this architecture that the rules do not have an externally observable state. Because the rules are triggered at the same time and their procedures go through without planning, this architecture is very fast, in contrast to other planning-based approaches. The main advantage of the reactive model is its speed. The robot can react immediately to unforeseen events, while other architectures do not even register them or have to plan how to deal with this event. The disadvantage is that the robot does not learn in this architecture. He does not create an internal world model and cannot plan optimal routes, let alone long-term strategies. This is a hindrance in areas of application in which only limited communication with the robot is possible.

Subsumption architecture

The subsumption architecture was introduced by Rodney Brooks in 1986. The model is based on the division of an overall behavior for the robot into many individual components ( modules ). Each module contains a behavior for a specific task, such as “following the wall” or “ collision avoidance ”. The modules are divided into hierarchical layers, which represent different levels of competence. Behaviors at higher competence levels can access inputs and outputs from modules at lower levels in order to create new behaviors from them. An example of this is a module “Move safely to coordinate X, Y”. This module could build on two further modules in the next lower layer, for example “Avoid collision” and “Move to X, Y”. Modules can influence other modules in lower layers. There are two types of influence, suppression and inhibition . Each behavior module has a control line. The suppression or inhibition signals can be sent to behavior further down the hierarchy via these control lines. When a module receives a suppression signal, the output of this module is replaced by the output of the module that sent the suppression signal. For example, the behavior of collision avoidance can be suppressed and instead the outputs of a “push-door-open” behavior can be sent to the motors. The inhibition signal behaves similarly, only that the output of the module to which it is sent is suppressed. In this way, a behavioral network is set up in which modules from lower layers are also retained, unlike in the hybrid model, where these are embedded in the reactive layer or deactivated as required. It has been shown that problems arise from this structure. For example, there is no need for a module that is suppressing another module to know what that other module is working on. Conversely, the suppressed module has no information about the module from which it is suppressed. In practice, however, there is often a need to exchange information between the individual layers. It is also problematic if a module in a higher hierarchy level suppresses a lower-level module, of which it does not know why it is performing its behavior. A module in layer 0, for example, has the option of recognizing abysses or steps and stopping them beforehand. If a module in layer 3 detects a socket and decides that the robot has to charge its batteries, the signal from layer 3 can suppress layer 0 and the robot could fall down stairs.

Function-oriented architecture (deliberative architecture)

The deliberative model (English deliberate - intentionally, intentionally) is in its properties the opposite of the reactive. The first step in a robot with a deliberative model is to get a picture of the world. So he creates a more or less accurate internal world model. In the second step, he plans his actions in this world model with a view to achieving his goals. The planned actions are now carried out. However, forward-looking planning requires an exact world model. The creation of such a world model cannot be exact and is always subject to errors. An example of this is an outdated plan. Suppose a robot has chosen a path to the next target that avoids obstacles such as stones. He now starts execution and begins moving. During the movement phase, an obstacle changes position - for example, rolling down a slope. Now the previously created world model is wrong and the robot could collide with this obstacle. In addition, the creation of the world model requires time, computing and storage capacity, which is not available in all environments, for example Mars robots. A prerequisite for the application of the deliberative model is that the robot knows all the causal consequences of its actions. Without such action-consequence pairs, it is not possible to build an action tree. Such an action tree is searched using search algorithms in order to obtain an action path to the desired destination. Fundamental for the execution of actions is of course the assumption that these actions run atomically up to their consequence (goal). By planning the actions, the deliberative model has advantages over the reactive. However, problems such as the outdated world model mentioned above also arise. Purely deliberative, i.e. serially working, robots can hardly be found in practice, as robots almost always run concurrent processes.

Hybrid architectures

The hybrid model tries to combine the advantages of the reactive and deliberative model. A typical hybrid model consists of three layers that only communicate with the adjacent layer. The first layer is a planning component that has the information about the planning objective and divides it into further sub-objectives (action packages). The planning component also has a world model that is necessary for deliberative planning. The second layer forms a link between the first and third layers. It is often called a sequencer in the literature. If necessary, the sequencer breaks down the action packages required by the planning component again and activates these packages or behavior in the third, reactive, shift. The sequencer also installs event monitors, which it can use to determine whether the desired behavior has taken place, whether unforeseen events have occurred or time frames have been exceeded. The third layer now contains all the necessary reactive behavior that is necessary to achieve the goal selected by the planning component. Only the third layer interacts directly and reactively with the environment. All three levels work asynchronously and with different time scales and data representations. The stratification mentioned is only one possible variant for the construction of a hybrid robot control architecture. Possible variations include the integration of the deliberative component (planning component) in the sequencer, or directly in the behavior.

The hybrid is the most widely discussed and used model in the literature. There are many variations on this architecture that incorporate elements of artificial intelligence such as neural networks. An example of this is the use of an Extended Kohonen Map for drive control. The advantage of the architecture lies in the merging of planning (deliberative) and execution (reactive), and their simultaneous execution. This enables reactive (fast) behavior and, at the same time, the calculation of a complete plan for the desired target state using methods of symbolic artificial intelligence. A disadvantage is the difficult coordination of the three levels of this model, due to their different data representations and time scales.

Cognitive architectures

In the case of cognitive architectures, the robot's ability to learn and the addition of new operational knowledge are in the foreground. The model for this are human cognitive processes. The aim is to model the dynamics, expandability and generalization of the system. Well-known approaches for this are SOAR and ACT-R .

References

Individual evidence

  1. ↑ spiegel.de: Drivers will be superfluous from 2018
  2. ^ RA Brooks: A robust layered control system for a mobile robot . In: IEEE Journal of Robotics and Automation . 2, No. 1, August, pp. 14-23. doi : 10.1109 / JRA.1986.1087032 .
  3. ^ JE Laird and A. Newell and PS Rosenbloom: SOAR: an architecture for general intelligence . In: Artificial Intelligence . 33, No. 1, September 1987, pp. 1-64.

literature

Web links