optimistic

This repo implements the code for the workshop paper Learning to Plan with Optimistic Action Models, which is a method for learning action models to be used within a task and motion planner (see paper for more details). This code performs experiments in a tool-use domain. It leverages PDDLStream for task and motion planning (TAMP), and pb_robot for integration with a Panda robot run in a pyBullet simulator.

Installation

This repo was developed using Python 3.9.6.

Clone this repo and install requirements

git clone git@github.com:carismoses/optimistic.git
cd optimistic
xargs -n1 python3 -m pip install < requirements.txt

Install pb_robot using the instructions below, not the ones in the repo's README. This installs dependecies, clones the repo and compliles the IKFask library for the panda (see Troubleshooting below if the build command fails)
```
cd ~
git clone git@github.com:carismoses/pb_robot.git
cd pb_robot/src/pb_robot/ikfast/franka_panda
python3 setup.py build
```
Install pddlstream using the instructions found there
Generate a symlink from pddlstream and pb_robot to this directory

cd ~/optimistic
ln -s ~/pddlstream/pddlstream .  # this assumes you installed it to your home directory
ln -s ~/pb_robot/src/pb_robot .  # this assumes you installed it to your home directory

Run

Collect Data and Train Classifiers

This will iteratively have the robot take actions and train action models on the collected data. See paper for the details of each data collection method.

python3 -m experiments.collect_data_and_train --data-collection-mode <DCM> --exp-name <EN> --max-actions <MA> --vis

--vis is an optional argument that allows you to visualize the actions the robot is taking in pyBullet. <EN> (required) saves results to logs/experiments/<EN>. <MA> is an integer number of actions to have the robot perform and train on. <DCM> (required) can be any of the following data collection methods:

'random-actions': sample and execute random actions (Random Actions method in paper)
'random-goals-opt': sample random goals, plan to achieve them using the optimistic model, execute found plan. Use random rollouts when no plan is found (Random Goals method in paper)
'sequential-plans': sample and execute actions found using the Sequential method (see paper)
'sequential-goals': sample and execute actions found using the Sequential Goals method (see paper)

Evaluating Model Accuracy

evaluate/model_accuracy.py is used to calculate the accuracy of trained models. The all_models_path variable contains the paths to the experiments to be evaluated and the test_dataset_path is the path to the experiment to be used as the test dataset.

Planning with Learned Models

To plan with the learned action models we developed a skeleton-based planner (see paper for details). The code for using this planner is in evaluate/plan_success.py.

Troubleshooting

This may be useful for setting up the repo using a different Python version. If building pb_robot with python setup.py build failed with the following error:

./ikfast.h:41:10: fatal error: 'python3.6/Python.h' file not found

The compiler can't find the appropriate python header. The solution is to first locate the header:

$ find /usr/local/Cellar/ -name Python.h
/usr/local/Cellar//python/3.7.7/Frameworks/Python.framework/Versions/3.7/include/python3.7m/Python.h
/usr/local/Cellar//python@3.8/3.8.2/Frameworks/Python.framework/Versions/3.8/include/python3.8/Python.h

which prints the python include directories. I wanted to use 3.7, so then I set the environment variable

export CPLUS_INCLUDE_PATH=/usr/local/Cellar//python/3.7.7/Frameworks/Python.framework/Versions/3.7/include/

and finally modify pb_robot/src/pb_robot/ikfast/ikfast.h by changing

#include "python3.6/Python.h" -> #include "python3.7m/Python.h"