Skip to content

Latest commit

 

History

History
166 lines (102 loc) · 8.06 KB

README.md

File metadata and controls

166 lines (102 loc) · 8.06 KB

Parameter Tuning

This package aims to implement the following two papers in researching global blackbox optimisation techniques for automated parameter tuning of motion planning algorithms.

Automatic Parameter Tuning of Motion Planning Algorithms (IROS 2018)

Automated Tuning and Configuration of Path Planning Algorithms (ICRA 2017)

Contents

Quick Start

Please run setup script if you did not initially setup with crane_plus_setup_FULL.sh. This script installs pip and all required submodules and dependencies to run the all parameter tuning modes.

$ chmod u+x parameter_tuning_setup.sh && ./parameter_tuning_setup.sh

Launch control.launch to:

  • Launch Gazebo simulation in tuning mode tuning:=true which sets up the physics properties in the tuning_world.world file to speed up realtime simulation
  • You can optionally set gui:=false to launch Gazebo without gui.
  • You can optionally set rviz:=false to launch Moveit without rviz.
$ roslaunch crane_plus_control.launch control.launch tuning:=true gui:=false rviz:=false

  1. Benchmarking session with OMPL or planner select defaults.

    $ roslaunch crane_plus_control parameter_tuning.launch planner_select:=Cano_etal mode:=default avg_runs:=5

  2. Parameter tuning session with selected mode.

    $ roslaunch crane_plus_control parameter_tuning.launch mode:=tpe max_runtime:=7200

  3. Benchmarking or tuning a specific path with defined start pose and target pose.

    $ roslaunch crane_plus_control parameter_tuning.launch mode:=ompl start_pose:=backbend target_pose:=low_fwd_reach

Note: Please ignore [WARN] Dropping first 1 trajectory point(s) out of 135, as they occur before the current time. First valid point will be reached in 0.130s. which outputs before every plan. This is a known issue that has been patched in newer updates of ros-kinetic-joint-trajectory-controller.

Parameter Tuning Parameters

Launch parameter tuning. Please feel free to adjust the defaults below.

$ roslaunch crane_plus_control parameter_tuning.launch

  • planner_select: Sets the desired planner configs as described in papers.

    • [default] Cano_etal
    • Burger_etal

  • planner: Sets the desired motion planner

    • [default] all
    • RRTConnect - bidirectional version of rapidly-exploring random tree (RRT)
    • BiTRRT - bidirectional version of RRT variant, transition-based RRT (TRRT)
    • BKPIECE - bidirectional version of KPIECE, a tree-based planner that uses discretisation to guide the exploration of the continuous space

  • mode: Sets the mode of the parameter tuning session.

    • ompl - Runs benchmarking session with OMPL planner config defaults.
    • [default] default - Runs benchmarking session with planner config defaults of specified planner select.
    • all - Runs tuning session session in all tuning modes.
    • tpe - Runs tuning session using Bayesian optimisation with TPE in Hyperopt.
    • rand - Runs tuning session using standard random search in Hyperopt.
    • smac - Runs tuning session using SMAC in SMAC3.
    • aucbandit - Runs tuning session using AUC Bandit in OpenTuner
    • gp - Runs tuning session using Bayesian optimisation with GP in skopt
    • rf - Runs tuning session using sequential optimisation with random forest regressor in skopt
    • et - Runs tuning session using sequential optimisation with extra trees regressor in skopt
    • gbrt - Runs tuning session using sequential optimisation with gradient boosting tree regressor in skopt

  • avg_runs: Sets the avg number of runs for each parameter configuration.

    • [default] 1

  • max_trials: Sets the max number of trials when in parameter tuning mode.

    • [default] 30

  • max_runtime: Sets the max runtime (secs) in parameter tuning mode. When set, overrides max_trials=10000.

    • [default] None

  • start_pose: Sets the start pose for specific path tuning.

    • [default] None

  • **target_pose: ** Sets the target pose for specific path tuning.

    • [default] None

OMPL Planning Time Benchmark

Adapted from moveit_benchmark_statistics.py from MoveIt! framework wih reference to benchmarking planners in OMPL. See here for more details on parameter input.

  1. Please run setup script if you did not initially setup with crane_plus_setup_FULL.sh.

    $ chmod u+x benchmark_db_setup.sh && ./benchmark_db_setup.sh

  2. Launch CRANE+V2 simulation. Adjust and save the desired scenes, queries and states to be benchmarked in the MongoDB Warehouse.

    $ roslaunch crane_plus_moveit_config demo.launch db:=true

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

  4. Save a scene on the Stored Scenes tab and name it Scene1 by double clicking the scene in the list.

  5. Move the start and goal states of the Crane arm by using the interactive markers.

  6. Save an associated query for the Scene1 scene and name the query Move1. Save a start state for the robot on the Stored States tab and name it Start1.

  7. The config file benchmark_config.yaml refers to the scenes, queries and start states used for benchmarking. Modify them appropriately.

  8. Set the username for the output_directory to export the benchmarked files. Log files will be saved to /home/${USER}/catkin_ws/src/crane_plus_v2_motion_planning/crane_plus_control/results/benchmarks/ by default.

    $ rosed crane_plus_control benchmark_config.yaml

  9. Bring down your previous demo.launch file ( Ctrl + C) and run the benchmarks.

    $ roslaunch crane_plus_control benchmark.launch

  10. Run moveit_benchmark_statistics.py to view results. A benchmark.db and benchmark_plots file will appear in the benchmarks folder. See --help for more options.

    $ rosrun crane_plus_control moveit_benchmark_statistics.py

  11. Click on the Change Database button to upload the benchmark.db file generated by script to plannerarena.org to interactively visualise results.