Traversability-based contact space planner code release. Given environment specified as a set of polygonal surface and a robot model, the code generate contact sequence for a humanoid robot. We provide an example with the Escher humanoid robot model. The code is written in Python 2.7 and tested in Ubuntu 14.04 with ROS Indigo.

• Install OpenRAVE, Installation Guide (Tested Commit: 7c5f5e27eec2b2ef10aa63fbc519a998c276f908)

• Install ROS

• Install srdfdom and tinyxml2: sudo apt-get install ros-indigo-srdfdom libtinyxml2-dev

• Install Eigen and NEWMAT: sudo apt-get install libeigen3-dev libnewmat10-dev

• Install cddlib (Tested Version: cddlib-094h):

  Get cddlib from ftp:

  wget ftp://ftp.math.ethz.ch/users/fukudak/cdd/cddlib-094h.tar.gz
  tar -xvf cddlib-094h.tar.gz
  cd cddlib-094h
  

  or from github:

  git clone https://github.com/cddlib/cddlib.git
  cd cddlib
  git checkout 0.94h
  

  Install cddlib and dependencies:

  sudo apt-get install libgmp3-dev
  ./configure
  make
  sudo make install
  cd /usr/local/include
  sudo mkdir cdd
  sudo mv cdd_f.h cddmp_f.h cddtypes_f.h cdd.h cddmp.h cddtypes.h setoper.h cdd
  

• Create a catkin workspace and clone the repo under /path/to/catkin/workspace/src, run catkin_make in /path/to/catkin/workspace to build the code.

• Before running the code, source /path/to/catkin/workspace/devel/setup.bash.

The contact planner is initiated with the script humanoid_motion_planner.py with the following options:

• contact_sequence_generation_method: Decide how contact sequence is generated for each guiding path segment.
  all_planning: Use graph search planning in every segment.
  all_retrival: Retrive and adapt previously generated motion plan in evey segment.
  hybrid(Default): Use planning in segment with high traversability, and use retrival in segment with low traversability.

• path_segmentation_type: Decide how guiding path is segmented.
  no_segmentation: The planner will use every motion mode along the guiding path.
  motion_mode_segmentation: The planner will segment the guiding path at where motion mode changes.
  motion_mode_and_traversability_segmentation(Default): The planner will segment the guiding path at where motion mode changes, and then further decompose each segment based on the traversability.

• traversability_threshold_type: The type of threshold that the planner use to determine the contact sequence generation method.
  mean(Default): The planner uses the mean of the traversability of all torso transition in a segment to determine using planning or retrival method to generate contact sequence in the segment.
  max: The planner uses the max of the traversability of all torso transition in a segment to determine using planning or retrival method to generate contact sequence in the segment.

• surface_source: The source of the environment in planning. The repo provides 3 examples each for two-corridor and two-stair environment. The user can create new environment by adding new options in update_environment function in environment_handler.py.
  two_corridor_environment(Default): Randomly generate a two-corridor environment.
  two_stair_environment: Randomly generate a two-stair environment.
  load_from_data: Load environment object file stored using pickle from path specified by environment_path parameter.

• environment_path: The folder which contains the stored environment object file.

• start_env_id and end_env_id: The first and last environment object file id loaded in the process.

Example Usage:

python humanoid_motion_planner.py contact_sequence_generation_method hybrid traversability_threshold_type mean path_segmentation_type motion_mode_and_traversability_segmentation environment_path environment_two_corridor surface_source load_from_data start_env_id 0 end_env_id 0

To swap a robot model, check the the comments in load_escher.py, and define all corresponding variables for the new robot model to create a robot loading function. Then swap all load_escher functions with the newly defined robot loading function.

To train traversability regressor, first generate the footstep_window, which precomputes all the footstep location given a torso transition. Then collect training data for each separate regressor mode (legs_only, legs_and_one_hand, all_manipulators). Finally, train the model with the training data. In summary, run the scripts with the following steps:

1. python transition_footstep_window_generator.py (Generate footstep_window)
2. python traversability_training_data_collector.py batch_id [training data batch id] mode [legs_only,legs_and_one_hand,all_manipulators] sample_env_num [number of sampling environment] surface_source [environment name] (Collect training data for each mode.)
3. python traversability_regressor_training.py mode [legs_only,legs_and_one_hand,all_manipulators] (Train traversability regressor for each mode.)