This package contains trajectory optimization algorithms meant predominantly for continuous control taks (simulated with MuJoCo).
The main package dependencies are MuJoCo and mjrl. See setup-instructions to get a working conda environment and setup dependencies.
After mujoco_py has been installed, the package can be used by either adding to path as:
export PYTHONPATH=<path/to/trajopt>$PYTHONPATHor through the pip install module
$ cd trajopt
$ pip install -e .The tricky part of the installation is likely to be mujoco_py. Please see instructions and known issues for help.
The algorithms assume an environment abstraction similar to OpenAI gym, but requires two additional functions to be able to run the algorithms provided here.
get_env_state() should return a dictionary with all the information required to reconstruct the scene and dynamics. For most use cases, this can just be the qpos and qvel. However, in some cases, additional information may be required to construct scene and dynamics. For example, in multi-goal RL, we can represent virtual goals using sites as opposed to real joints.set_env_state(state_dict) should take in a dictionary, and use the contents of the dictionary to recreate the scene specified by the dictionary.
The example reacher environment has an illustrative.See this directory for illustrative examples: trajopt/examples.
If you find the package useful, please cite the following paper.
@INPROCEEDINGS{Lowrey-ICLR-19,
AUTHOR = {Kendall Lowrey AND Aravind Rajeswaran AND Sham Kakade AND
Emanuel Todorov AND Igor Mordatch},
TITLE = "{Plan Online, Learn Offline: Efficient Learning and Exploration via Model-Based Control}",
BOOKTITLE = {ICLR},
YEAR = {2019},
}