MAGI (Multistep Action Graph Instantiator) lets you construct a multistep action graph and find a sequence of executable trajectories using various backtracking strategies and PrPy planners.

To install and run MAGI you should folow the development environment guide to create your Catkin workspace. Note that these instructions assume that you followed the machine setup guide on your computer or virtual machine.

In summary, you should be able to execute these commands to download and build the code:

$ cd my-workspace/src
$ touch .rosinstall
$ wstool set magi --git https://github.com/personalrobotics/magi -y
$ wstool merge https://raw.githubusercontent.com/personalrobotics/pr-rosinstalls/master/herb-minimal-sim.rosinstall
$ wstool set pr-ordata --git https://github.com/personalrobotics/pr-ordata -y
$ wstool update
$ cd ..
$ rosdep update
$ find src -name manifest.xml -delete
$ rosdep install -y -r --ignore-src --rosdistro=indigo --from-paths src
$ catkin build
$ . devel/setup.bash

To run an example script, start roscore in a separate terminal and try

$ rosrun magi grasp_glass_demo.py --viewer rviz

Action defines what needs to be planned. Each Action has plan method through which MAGI generates a Solution, which contains geometrically feasible path that can be postprocessed to be executed on the robot. MAGI offers many actions, including the following:

• MoveHandAction: Moves hand DOFs. OpenHandAction, CloseHandAction, GrabObjectAction, ReleaseObjectAction inherit from this.
• PlaceObjectAction: Places an object onto another object.
• PlanAction: Uses PrPy's PlanTo... methods to generate paths. PlanToTSRAction, PlanEndEffectorStraight inherit from this.
• MoveUntilTouchAction: Uses PrPy planner to move along a direction until the endeffector touches an object.

Actions can be combined to create a multistep action. MAGI has several meta-actions.

• SequenceAction takes a list of Action and creates one action sequence.
• ParallelAction takes a list of Action and creates a tree where each Action is a branching node.
• DisableAction: Disables links of an object while planning an Action. Useful when the wrapped Action results in collision with the padding of an object, which many PrPy planners will not succeed unless the padding is disabled.

An action graph is one meta-action in which all necessary sequences and branches of actions are defined. Use a combination of Action and meta-actions to construct an action graph for the multistep task you intend to perform. For example, an action graph for moving an object can be written as a SequenceAction as the following:

actions = [
    OpenHandAction(...),

    # Plan to somewhere near object
    PlanToTSRAction(...),

    # Move until the robot touches the object
    MoveUntilTouchAction(...)

    GrabObjectAction(...)
]

SequenceAction(actions, name="GrabObjectSequence")

MAGI offers two backtracking strategies to search through an action graph.

• DepthFirstPlanner tries each non-deterministic node multiple times before backtracking.
• RestartPlanner restarts from the root whenever an action node fails.

If it succeeds, Planner returns a solution which can be executed.

planner = RestartPlanner()
with env:
    solution = planner.plan_action(env, action)
execute_pipeline(env, solution)