[Documentation] [TitleIndex] [WordIndex

The lama_jockeys package provides base class for Learning, Localizing, Navigating, and custom jockeys for the Large Maps Framework (LaMa). A jockey provides a specific capability to the robot. These modules are called jockeys since, in a certain sense, they are riding the robot. It means that these modules have direct access to the robot hardware and are able to control the robot's actuators, to read its sensors and to process raw data gathered from its sensors.

The jockeys are actionlib servers. The lama_jockeys packages provides standard action messages for Learning, Localizing, and Navigating jockeys.

A navigating jockey takes care of traversing edges with respect to a specific procedural knowledge. It guides the robot along the given edge and determines the end of this edge. If it is possible to interrupt the execution of the navigating algorithm, the navigating jockey is call interruptible. This property allows to restart it later from a slightly different position while the algorithm still keeps the capability to navigate to the original target vertex.

A navigating Jockey has two different subtypes: memory-less and memory-based.

Memory-less Navigating Jockey

A memory-less jockey uses a reactive navigation strategy. The procedural knowledge represented by a reactive navigating strategy has a small or inexistent use of database data. Only the information needed to distinguish outgoing edges from each other is stored in the map.

The reactive algorithm computes the control commands directly from the sensor input and has no persistent internal state. Therefore, a memory-less navigating jockeys are often interruptible.

The main advantage of reactive navigation is the ability to traverse an unknown edge. It allows to discover new locations, which are the target vertex of these edges. Therefore the memory-less Jockeys are used mainly during the exploration of the environment.

Memory-based Navigating Jockey

A memory-based jockey needs the data part of the procedural knowledge about the edge for proper operation. The data for the memory-based jockeys are typically more complex than for the memory-less ones. These data (edge descriptor) are stored in the map and must be inserted into the map by a coupled learning jockey in advance. Therefore, a memory-based navigating jockey can guide the robot along previously visited edges only.

A memory-based navigation should be more precise and repeatable than a reactive one. Such jockey can determine the failure of the navigation and report it to the higher-level control module. In addition, the jockey can estimate in advance the possibility of failure according to the given edge descriptor.

Moreover, such jockey can provide the measurement of similarity of appropriate edges.

Learning Jockeys

A learning jockey is usually tightly coupled with a particular memory-based navigating jockey which utilizes the procedural knowledge of the same type. Such a jockey produces the procedural knowledge by gathering the information while a robot can be driven by another navigating jockey (typically memory-less), as a learning jockey is not supposed to control the robot actuators while traversing an edge.

The gathered procedural knowledge is divided into algorithmic and data parts. The data part is stored in the map as edge descriptor. This edge descriptor can be utilized by a memory-based navigating Jockey later on.

It is possible, that a learning jockey needs to control the robot actuator occasionally, in order to increase the performance, get more data from the sensor, get more time for data processing, or handle errors. In this case, the learning jockey can request the high-level control module to interrupt the running navigating jockey. If the navigating jockey is interruptible, the high-level control module suspends the navigating jockey and grants the control to the learning jockey. The original navigating jockey is reactivated as soon as the learning jockey achieves its goal and notifies the high-level controller that the reason for interruption disappeared.

A learning jockey extends an existing edge with new procedural knowledge, which can then be used to traverse the edge. This property allows to build rich but unified spatial representation for different types of robots equipped with wide range of sensors. Such a map is suitable for a heterogeneous robotic team, as it allows easy sharing of the spatial knowledge among team members.

Localizing Jockeys

Localizing jockey are associated to specific types of vertex descriptors and specific sensors. Such jockeys process the sensor data into a vertex descriptor. The resulting descriptor is stored in the map using the lama_interfaces. Different sensors require different localizing jockeys for handling specific hardware and processing data, nevertheless these jockeys can utilize the same map interface and therefore share the stored information. On the other hand, localizing jockeys with different algorithms can share the same sensors.

When the robot arrives to a vertex, the jockey gathers information about the surrounding environment, computes and stores specific descriptors into the map and possibly discovers outgoing edges from the current vertex. Different localizing jockeys can produce different vertex descriptions and also discover different outgoing edges. A localizing jockey is able to distinguish vertices from each other and compare the actual vertex with other vertices already stored in the map. The result of its action is to give the probabilities of being in some of the previously visited vertices. This behavior is used for global localization and loop-closing.

In addition, a localizing jockey can estimate the robot relative position in the vertex local frame of reference if the stored descriptor has the form of a local metric map. This ability is crucial if the robot has to interact with the environment.

Custom Jockeys

In the case that the functionalities of a jockey cannot be described by the predefined actions Navigate.action, Learn.action, and Localize.action, the jockey can be based on the lama_jockeys::Jockey class instead. Basic map and jockey functionalities are already implemented.


The lama_jockeys::Jockey provides the mechanisms to interfere with the map in the form of the member variable map_agent_. For example, to get the vertex with id 243 from the map, one can write:

   1 #include <lama_jockeys/jockey.h>
   3 lama_interfaces::ActOnMap srv;
   4 srv.request.action = lama_interfaces::ActOnMapRequest::PULL_VERTEX;
   5 srv.request.object.id = 243;
   6 ROS_DEBUG("Calling action PULL_VERTEX for vertex %d", 243);
   7 if (!map_agent_.call(srv))
   8 {
   9   ROS_ERROR("Failed to call map agent");
  10   /* throw an error or abort the action: "server_.setAborted();" */
  11   return;
  12 }
  13 ROS_DEBUG("Received response PULLL_VERTEX");
  14 lama_interfaces::LamaObject vertex = srv.response.objects[0];

The three subclasses lama_jockeys::LearningJockey, lama_jockeys::LocalizingJockey, and lama_jockeys::NavigattingJockey handle reception of the goal, which is then stored in the goal_ member variable. Upon goal reception, one specific function is called. Most of these functions are implemented are pure virtual method, so that the jockey implementer must only define these methods to have a working jockey. A default implementation is provided for the common actions "INTERRUPT" ("onInterrupt()") and "CONTINUE" ("onContinue()"), they only manage to pause the stop watch which is otherwise running when an action is running.

Learrning Jockeys

The user must implement "onStartLearn()" and "onStopLearn()".

Localizing Jockeys

The user must implement "onGetVertexDescriptor()", "onGetEdgesDescriptors()", "onLocalizeInVertex()", "onLocalizeEdge()", "onGetDissimilarity()".

The user must implement "onTraverse()" and "onStop()".

2024-07-13 13:17