knowledge_representation  master
Persistent knowledge for robots


Mechanisms for storing and querying information about the world. A nicely packaged database schema to structure the kinds of knowledge that a robot might need and tools for CRUD operations within that structure. Integrate it into your robot's perception so you can query where objects were seen, or into your task planning system so you can easily manage all of your facts.

  • Persistent storage of facts and map geometry backed by a local PostgreSQL database
  • Convenient APIs for common queries, like getting all instances of a type of object
  • Link directly against your C++ code, or import a Python binding
  • Not dependent on ROS, but easy to integrate into ROS projects


Knowledge is broken up into concepts, instances and attributes.

Concepts are abstract ideas. They represent the idea of something. For example, the general idea of an apple but not any one apple in particular.

Instances are concrete, usually physical manifestations of concepts. You could have a dozen instances of a concept like apple.

Both concepts and instances are entities in the knowledgebase. Each entity is granted a unique identifying number.

Attributes tie entities together and store data about them. Each attribute has a name and a type. For example, an apple instance might be on some other instance. The attribute is named is_on and its type is entity_id, because it relates some entity to another. Attributes with a value type of entity_id are also called relations because they relate entities to each other. Attributes with types other than entity_id function more like properties; for example, the relation is_graspable with a type of bool is just storing a simple fact about an entity.

Map Types

Mobile robots often have to answer questions like "what room am I in?" or "which is the nearest X?". knowledge_representation includes special affordances for representing 2D geometric information that can be queried to answer these kinds of questions:

  • Points are a special kind of instance which store an x and y coordinate.
  • Poses are a special kind of instance which store an x and y coordinate as well as a direction.
  • Regions are a special kind of instance which store a list of x and y coordinates defining a closed region.
  • Doors are a special kind of instance which store a pair of x and y coordinates defining a door.

All of these types are uniquely tied to a single map, a special kind of instance.

Like concepts, all of these geometric types must have names configured. Without names attached, there isn't enough semantic information to support any kind of useful operation. What would it mean for a point to be in a region if neither the point nor the region had names?


Make sure you've pulled down the package dependencies using rosdep

rosdep install --from-paths . -ry

Then run one of scripts/configure_{mysql,postgresql}.sh to install the default database configuration and schema. PostgreSQL is the preferred backend.


Integrate knowledge_representation by using the API wherever you need to store facts and observations that the robot might need later. Use the same API to retrieve facts en masse so you can plan over them, inspect them, or do whatever else you need for your application.

C++/Python API

To create applications that store and query, link your C++ against the knowledge_rep library or just import the knowledge_representation Python module. See the documentation for the latest version of the C++ API and example usage in test/*.cpp. The Python API is a generated wrapper, so must classes and methods are the same but with snake case conventions. See scripts test_ltmc or scripts/show_me for example usage, and try out the ikr script to interactively explore the API.

Loading Knowledge

Scripts are provided for bulk loading knowledge. This is helpful if say, you want to load in map annotations, or you have an ontology that you want to use to initialize the robot's knowledge.

populate_with_map loads in SVG map annotations (marked over an image, ROS-style map). See test/resources for some example SVG files. In the future, we'll provide a simple browser-based annotator as well.

populate_with_[knowledge|owl|xml] support loading in different kinds of ontologies. Documentation and example files will come in a later release.


Once your robot has accumulated knowledge, you'll want to poke around. Use the show_me script to quickly see a summary of the current knowledge, then pass it an ID or a name to see details about entities and their relations.

As Part of a ROS System

Once you add knowledge_representation to your package's dependencies, your code will link against the library that provides the APIs.

The populate_ scripts are geared towards use in a launch files. For instance, to load a map on launch, just add a line like this to your launch file:

<node name="load_map_annotations" pkg="knowledge_representation" type="populate_with_map" args="$(find my_map_package)/maps/map_name.yaml"/>

Once the package is built in a catkin workspace, you can access additional shell shortcuts:

  • ikr: run the iPython environment
  • kr-query: run a SQL query against the knowledgebase
  • kr-save: save a SQL-dump of the current knowledgebase
  • kr-show: run show_me


If you're working on the MySQL interface, we access the backing store via the xdev API. See the documentation for the official MySQL xdev API C++ library.