ROS related frequently asked questions.
The FAQ from the py_trees package.
You can imagine once you have 50+ re-usable behaviours in a tree that the need for remapping of topics, services and parameters in the behaviour tree launch description will become exceedingly large. In these situations it is more convenient to load parameters for these remappings in a structured way on the parameter server (loaded from a single yaml). This centralises your application configuration and additionally exposes that configuration at runtime which will assist with debugging. Sanity…
On the parameter server, such configuration might look like:
/tree/topics/odom /gopher/odom /tree/topics/pose /gopher/pose /tree/services/get_global_costmap /move_base/global/get_costmap /tree/parameters/max_speed /trajectory_controller/max_speed
In code, highlighting re-usability of the remappings across multiple behaviours:
odometry_topic=self.node.get_parameter_or(name="~topics/odom", alternative_value="/odom") pose_topic=self.node.get_parameter(name="~topics/pose", alternative_value="/pose") move_base = my_behaviours.MoveBaseClient(odometry_topic, pose_topic) odometry_foo = my_behvaiours.OdometryFoo(odometry_topic)
Even though the behaviour tree provides a continuous tick-tock method, you can set your own pace. This can be useful if you wish to vary the tick duration, or to tick only when an external trigger is received (a common trick to minimise cpu usage in games). For example:
... while rclpy.ok(): rclpy.spin_once(timeout_sec=0.1) if some_external_trigger: tree.tick_once()
Triggering based on logic inside the tree however, is much more challenging as this is almost a chicken and egg situation (tick only when an event fires, but events are typically embedded in the decision making tree itself). UE4 has an implementation that has crafted mechanisms for this which go beyond basic behaviour tree concepts - if you have such a need, it’s likely you’ll have to extend py_trees to meet the needs of your own use case.
Control-Level Decision Making¶
Our first use case never intended to utilise behaviour trees for decision making typically considered internal to control subsystems. A good example of such is the approach logic for a docking maneuvre. Another is the recovery behaviours for navigation, which start to access sound/light notifications, specialised sensing contexts as well as specialised maneuvres. Note that neither of these require low-latency for their decision logic. Nonetheless, it was surprising to find the control engineers moving the logic from internal state machines to the behaviour trees at a higher level.
In hindsight, this makes good sense. With the robot’s decision making logic landing in one place, logging, debugging and visualising the state of the robot became simpler and could make use of a single set of tools. A growing library of shared and reusable patterns sped up the development cycle. It also liberated subsystems from having to co-ordinate other subsystems (e.g. the navigation system when engaging in recovery behaviours).