|Title:||Diagnostic System for Robots Running ROS|
|Author:||Tully Foote <tfoote at willowgarage.com>|
Monitoring and characterizing the functional state of a robot is important at all times. To be able to know this across multiple robot types and versions this REP lays out a standard way to report system diagnostics. These diagnostics provide three capabilities, a quick glance method of knowing that all systems are operating nominally, a way to access detailed information for debugging, and a long term logging infrastructure for historical analysis.
When operating a robot being aware that all the parts are running correctly is important. To accomplish this there must be a consisten way to present the data to the user. In this interface they need to be able to see a high level summary. But in the case that they need details they need to be able to quickly see fine grained details of any specific component.
In addition to realtime viewing collecting historical data for offline analysis is also important. This can be valuable to identify trends in data which can warn of future failures as well as help debug a failure which happened in the past but was not recognized until later.
To be useful the diagnostic system must be flexible enough to work on different robots, in different configurations without changing the user experience.
The core of the diagnostic system is the reporting mechanism. The goal of providing a generic reporting mechanism means that the diagnostics uses weakly typed information unlike the rest of ROS because the intended consumer is a person. These data structures are designed for aggregation and presentation to a user to be able to quickly check that all systems are running correctly while suppressing the details. And in the case they are not the user can quickly drill down and see all the details of any specific component.
For every component in the system the following data will be published:
This is the name of the component or subcomponent in the system. The component name is usually the node name of the driver node. For example if the node prosilica_driver is running it would report as:
It may be the case that subcomponents should be split out if they have enough information to stand on their own and could be interpreted independently of other subcomponets. For example if a driver is providing multiple topics one subcomponent per topic can be useful, since this is created automatically by helpers in diagnostic_updater, and one for the node as a whole. This avoids mixing in detailed timing statistics for the topics with core information about the hardware. If publishing for a subcomponent the component name should be component name: subcomponent name It might look like these to example subcomponets:
prosilica_driver: Packet Status prosilica_driver: Frequency Status
Another use case is if a driver provides an interface for multiple discrete components, such that providing seperate component names can be a useful distinction for the user.
These levels should be used to color GUI viewers with the designers equivilant of 'green', 'yellow', and 'red' symbols.
A human readable summary of the status of the device. This is often a concatenation of any error messages or a default message.
If applicable this identifies the specific hardware running. This is for things like serial numbers for devices so that a piece of hardware can be tracked between robots if it is moved between robots or moved within a robot.
As there is are uncountable types of hardware potentially added to the system each with their own specific pertinent information the hardware specific diagnostics data is captured in string key value pairs.
Common data to be published are settings, serial numbers, firmware versions, error counts, and information on latest errors or timeouts.
Reporting is carried out by message publication on the topic /diagnostics using the diagnostic_msgs/DiagnosticArray data type. The default publication rate is 1Hz.
Diagnostic outputs should be enabled for hardware components which are expect to be present whenever the robot is running.
The system is generic enough to handle all software components, however adding diagnostics to all software components creates too much noise for the core functionality to work consistently. Setting up the analyzers to make sure that important messages are shown and there are not false positives of important errors is impossible for all users use cases. Also if all pieces of software log to the diagnostics system the burden of logging and analyzing the logs goes up significantly.
There are possible solutions such as seperate topics for different types of diagnostics, however as this REP is targeted for hardware diagnostics only hardware based diagnostics sould be published over this protocol.
The diagnostic system has been designed to provide the operator with awareness of the current state of the system as well as provide a history of the state of the system for historical analysis.
Whenever a robot is operating the operator should have an instance of robot_monitor visible on a screen. This may be contained withing another app like pr2_dashboard. This will provide good situational awareness for the operator.
In the default launchfile used to bring up the hardware there should be a rosbag record instance setup to recored the /diagnostics topic, and periodically uploaded off the robot. For example:<!-- Runtime Diagnostics Logging --> <node name="runtime_logger" machine="c1" pkg="rosbag" type="record" args="-O /hwlog/pr2_diagnostics /diagnostics --split=2000" />
- This is not designed to be a keepalive, it uses potentially unreliable transports and does not have tight timeouts, and there may be stale data due to aggregation.
- This is not going to halt the system in any way. If there is an unsafe condition it must be dealt with independently. (For example on the PR2 in the case of a motor error, all motors halt, in addition to sending an Error diagnostic message. ) The diagnostic message is for operator awareness.
While user-end tools are not needed to generate and capture the diagnostic information, they perform a critical role in making the captured data accessible for analysis as well as making implementations of diagnostics much easier. More documentation can be found in the diagnostics stack .
When displaying diagnostic data there often some analysis to make data useful for a specific application. The diagnostic_aggregator is designed to do just this. It aggregates the latest information from each component and passes it to a configurable set of analyzers. The analyzers are useful for doing things like grouping outputs, suppressing outputs which are invalid for a specific application or configuration. diagnostic_aggregator wiki page 
Being able to quickly understand the status of all components in a system is important, and to do so a concise visualization tool was developed. When use with the aggregator above it will pop up all warnings and errors to the top level as well as providing a higherarchical view of the system layed out by the aggregators. robot_monitor wiki page 
# This message is used to send diagnostic information about the state of the robot Header header #for timestamp DiagnosticStatus status # an array of components being reported on
# This message holds the status of an individual component of the robot. # # Possible levels of operations byte OK=0 byte WARN=1 byte ERROR=2 byte level # level of operation enumerated above string name # a description of the test/component reporting string message # a description of the status string hardware_id # a hardware unique string KeyValue values # an array of values associated with the status
string key # what to label this value when viewing string value # a value to track over time
This document has been placed in the public domain.