Benchmarking

Getting Started

If you haven’t already done so, make sure you’ve completed the steps in Getting Started.

The benchmarking package provides methods to benchmark motion planning algorithms and aggregate/plot statistics using the OMPL Planner Arena. The example below demonstrates how the benchmarking can be run for a Panda robot arm.

Example

An example is provided in the examples folder. The launch file requires a MoveIt configuration package for the Panda robot arm available from here.

To run:

  1. Launch the panda_moveit_config demo.launch:

    roslaunch panda_moveit_config demo.launch db:=true

  2. Within the Motion Planning RViz plugin, connect to the database by pressing the Connect button in the Context tab.

  3. Save a scene on the Stored Scenes tab and name it Kitchen1 by double clicking the scene in the list.

  4. Move the start and goal states of the Panda arm by using the interactive markers.

  5. Save an associated query for the Kitchen1 scene and name the query Pick1.

  6. Also save a start state for the robot on the Stored States tab and name it Start1.

  7. The config file moveit_ros/benchmarks/examples/demo1.yaml refers to the scenes, queries and start states used for benchmarking. Modify them appropriately.

  8. Bring down your previous launch file (ctrl+c).

  9. Change the location output_directory to export the benchmarked files:

    rosed moveit_ros_benchmarks demo1.yaml

  10. Run the benchmarks:

    roslaunch moveit_ros_benchmarks demo_panda.launch

Viewing Results

The benchmarks are executed and many interesting parameters are aggregated and written to a log file. A script (moveit_benchmark_statistics.py) is supplied to parse this data and plot the statistics.

Run:

rosrun moveit_ros_benchmarks moveit_benchmark_statistics.py <path_of_logfile>

To generate a PDF of plots:

python moveit_benchmark_statistics.py -p <plot_filename> <path_of_logfile>

Alternatively, upload the database file generated by moveit_benchmark_statistics.py to plannerarena.org and interactively visualize the results.

Parameters of the BenchmarkOptions Class

This class reads in parameters and options for the benchmarks to run from the ROS parameter server. The format of the parameters is assumed to be in the following form:

benchmark_config:

  warehouse:
      host: [hostname/IP address of ROS Warehouse node]                            # Default localhost
      port: [port number of ROS Warehouse node]                                    # Default 33829
      scene_name: [Name of the planning scene to use for benchmarks]               # REQUIRED

  parameters:
      runs: [Number of runs for each planning algorithm on each request]           # Default 10
      group: [The name of the group to plan]                                       # REQUIRED
      timeout: [The maximum time for a single run; seconds]                        # Default 10.0
      output_directory: [The directory to write the output to]                     # Default is current working directory

      start_states: [Regex for the stored start states in the warehouse to try]    # Default ""
      path_constraints: [Regex for the path constraints to benchmark]              # Default ""
      queries: [Regex for the motion plan queries in the warehouse to try]         # Default .*
      goal_constraints: [Regex for the goal constraints to benchmark]              # Default ""
      trajectory_constraints: [Regex for the trajectory constraints to benchmark]  # Default ""

      workspace: [Bounds of the workspace the robot plans in.  This is an AABB]    # Optional
          frame_id: [The frame the workspace parameters are specified in]
          min_corner: [Coordinates of the minimum corner of the AABB]
              x: [x-value]
              y: [y-value]
              z: [z-value]
          max_corner: [Coordinates of the maximum corner of the AABB]
              x: [x-value]
              y: [y-value]
              z: [z-value]

  planning_pipelines:
      - name: [Name of the planning pipeline used as relative parameter namespace] # REQUIRED
        planners:                                                                  # REQUIRED
          - A list of planners
          - available in the
          - planning pipeline
          - to benchmark the
          - queries in.
      - name: ...
          - ...

Parameters of the BenchmarkExecutor Class

This class creates a set of MotionPlanRequests that respect the parameters given in the supplied instance of BenchmarkOptions and then executes the requests on each of the planners specified. From the BenchmarkOptions, queries, goal_constraints, and trajectory_constraints are treated as separate queries. If a set of start_states is specified, each query, goal_constraint, and trajectory_constraint is attempted with each start state (existing start states from a query are ignored). Similarly, the (optional) set of path constraints is combined combinatorially with the start query and start goal_constraint pairs (existing path_constraint from a query are ignored). The workspace, if specified, overrides any existing workspace parameters.

The benchmarking pipeline does not utilize MoveGroup. Instead, the planning pipelines are initialized and run directly including all specified PlanningRequestAdapters. This is especially useful for benchmarking the effects of smoothing adapters.

It is possible to customize a benchmark run by deriving a class from BenchmarkExecutor and overriding one or more of the virtual functions. For instance, overriding the functions initializeBenchmarks() or loadBenchmarkQueryData() allows to specify the benchmark queries directly and to provide a custom planning scene without using ROS warehouse. An example is the custom benchmark CombinePredefinedPosesBenchmark which expects a list of predefined joint states and then creates queries for all pair-wise combinations. See the config file demo_panda_predefined_poses.yaml for how to configure the poses. You can run this example with:

roslaunch moveit_ros_benchmarks demo_panda_predefined_poses.launch

Additionally, a set of functions exists for ease of customization in derived classes:

  • preRunEvent: invoked immediately before each call to solve

  • postRunEvent: invoked immediately after each call to solve

  • plannerSwitchEvent: invoked when the planner changes during benchmarking

  • querySwitchEvent: invoked before a new benchmark problem begin execution

Note, in the above, a benchmark is a concrete instance of a PlanningScene, start state, goal constraints / trajectory_constraints, and (optionally) path_constraints. A run is one attempt by a specific planner to solve the benchmark.

Benchmarking of Different Motion Planners: CHOMP, STOMP and OMPL

This section contains the instructions for benchmarking different motion planners present in MoveIt: CHOMP, STOMP and OMPL. These planners can be compared with each other for a well defined benchmark (which is for the same environment, start states, queries and goal states). Different metrics for each of the planners can be reported to get quantitative statistics which could aid in proper selection of a particular planner in a defined environment. The statistics reported for each of the planners includes: time taken to compute the path, path length, path time, whether a valid path was found or not, etc.

Benchmarking in a scene without obstacles

To benchmark different planners in a simple environment without obstacles, open one terminal and follow the steps in the Example section at the top of this page. In the last two steps instead of editing demo1.yaml and launching demo_panda.launch, edit demo_panda_all_planners.yaml and launch demo_panda_all_planners.launch.

Benchmarking in a scene with obstacles

To benchmark motion planners in a scene filled with obstacles, open two terminals. In the first terminal start RViz and wait for everything to finish loading:

roslaunch panda_moveit_config demo.launch db:=true

In the second terminal, run either of the two commands:

rosrun moveit_tutorials collision_scene_example.py sparse

or:

rosrun moveit_tutorials collision_scene_example.py cluttered

Now follow these steps:

  1. Within the Motion Planning RViz plugin, connect to the database by pressing the Connect button in the Context tab.

  2. Save a scene on the Stored Scenes tab and name it ObstaclesScene by double clicking the scene in the list.

  3. Move the start and goal states of the Panda arm by using the interactive markers.

  4. Save an associated query for the ObstaclesScene scene and name the query Pick1.

  5. Also save a start state for the robot on the Stored States tab and name it Start1.

  6. The config file moveit_ros/benchmarks/examples/demo_obstacles.yaml refers to the scenes, queries and start states used for benchmarking. Modify them appropriately.

  7. Bring down your previous launch file (ctrl+c).

  8. Change the location output_directory to export the benchmarked files:

    rosed moveit_ros_benchmarks demo_obstacles.yaml

  9. Run the benchmarks:

    roslaunch moveit_ros_benchmarks demo_panda_all_planners_obstacles.launch

To view the results follow the same steps as listed in the Viewing Results section above. After loading the database into Planner arena, different statistics can be analysed about each of the planners by choosing the required benchmark attribute from the drop down list (see figure). See image below for analysis of time taken by each of the planners to compute the solution for a sample benchmark.

../../_images/planners_benchmark.png

开源反馈

有地方需要改进？请在 GitHub page 新开一个 pull request。