Description

ldd_plan_node performs kinematic planning for the PR2 arm using the LDD algorithm, documented here:

Paul Vernaza and Daniel D. Lee. Learning dimensional descent planning for a highly-articulated robot arm. In Proceedings of the International Conference on Intelligent Robots and Systems, 2011.

Usage

The planner runs as a ROS node that serves GetMotionPlan requests. It may be launched by executing roslaunch ldd_plan_node ldd_plan_node.launch.

Various environment variables can be set that affect the performance of the algorithm. A convenient way to experiment with different parameters is to prepend them to the launch command as so:

option1=true option2=100 roslaunch ldd_plan_node ldd_plan_node.launch

Parameters

The following parameters may be set via environment variables:

* bboxMargin (float > 0, default 1.0)

• In each iteration, LDD searches for variations of the current path in some direction. The larger this parameter, the further LDD searches along each direction. If LDD cannot find a feasible path for your problem, increasing this parameter may help. Visualizing the paths with plotting=true may help to determine whether this parameter needs to be increased.

* nPath0 (int > 0, default 100)

• LDD discretizes the initial path (a straight line) by choosing nPath0 samples along its length. Suppose that the samples chosen in this way are a distance L apart. In subsequent iterations, the path is resampled so that the distance between samples is exactly L. Increasing this parameter therefore increases the resolution of the search.

* nGradSamp (int > 0, default 10000)

• LDD will draw this many random samples of the cost gradient in order to learn a basis for dimensional descent. Increasing this parameter will decrease the variance in the paths found by LDD, at the expense of increased runtime.

* nCostThreads (int > 0, default 1)

• LDD will spawn this many threads in parallel to compute the cost function and cost gradients.

* minPlanTime (float > 0, default 0.)

* maxPlanTime (float > 0, default infinity)

• LDD will optimize the path for at least minPlanTime seconds. It will then return the best path found so far, if a feasible path has been found, or continue planning until a feasible path is found or maxPlanTime is reached (whichever comes first).

* minIt (int > 0, default 0)

* maxIt (int > 0)

• The minimum and maximum number of iterations to perform before termination.

* dimMaxOpt (int >= 0, default -1)

• LDD cycles through a set of variation directions, optimizing the path sequentially in each of these directions, in order of "decreasing importance" before looping around. dimMaxOpt is the index of the maximum dimension to be optimized before looping around. If set to -1, LDD will optimize all dimensions.

* fCheckSampRate (float > 0, default 20.)

• At each iteration, the current path is resampled at a rate given by fCheckSampRate samples per radian and then checked for feasibility. The lower this parameter, the more likely it is that ldd_plan_node will consider feasible a path that move_arm considers infeasible (and will hence abort). Setting this parameter too high, however, will adversely affect performance.

* plotting (boolean, default false)

• If set to true, a 2D visualization will be shown illustrating the cost restricted to the currently-optimized submanifold and the optimal path found restricted to this sub manifold.

Known issues

* Important parameters in ldd_plan_node.launch are currently undocumented

* right_arm is currently hard-coded in LDDPlan.hh