mbrl is a toolbox for facilitating development of Model-Based Reinforcement Learning algorithms. It provides easily interchangeable modeling and planning components, and a set of utility functions that allow writing model-based RL algorithms with only a few lines of code.
See also our companion paper.
mbrl requires Python 3.7+ library and PyTorch (>= 1.7).
To install the latest stable version, run
pip install mbrl
If you are interested in modifying the library, clone the repository and set up
a development environment as follows
git clone https://github.com/facebookresearch/mbrl-lib.git pip install -e ".[dev]"
And test it by running the following from the root folder of the repository
python -m pytest tests/core python -m pytest tests/algorithms
As a starting point, check out our tutorial notebook
on how to write the PETS algorithm
(Chua et al., NeurIPS 2018)
using our toolbox, and running it on a continuous version of the cartpole
Provided algorithm implementations
MBRL-Lib provides implementations of popular MBRL algorithms
as examples of how to use this library. You can find them in the
mbrl/algorithms folder. Currently, we have implemented
PETS and MBPO, and
we plan to keep increasing this list in the near future.
The implementations rely on Hydra
to handle configuration. You can see the configuration files in
environment specific configurations for each environment, overriding the
default configurations with the best hyperparameter values we have found so far
for each combination of algorithm and environment. You can run training
by passing the desired override option via command line.
For example, to run MBPO on the gym version of HalfCheetah, you should call
python -m mbrl.examples.main algorithm=mbpo overrides=mbpo_halfcheetah
By default, all algorithms will save results in a csv file called
inside a folder whose path looks like
you can change the root directory (
./exp) by passing
root_dir=path-to-your-dir, and the experiment sub-folder (
experiment=your-name. The logger will also save a file called
model_train.csv with training information for the dynamics model.
Beyond the override defaults, You can also change other configuration options,
such as the type of dynamics model
dynamics_model=basic_ensemble), or the number of models in the ensemble
dynamics_model.model.ensemble_size=some-number). To learn more about
all the available options, take a look at the provided
Running the provided examples requires Mujoco, but
you can try out the library components (and algorithms) on other environments
by creating your own entry script and Hydra configuration (see [examples].
If you do have a working Mujoco installation (and license), you can check
that it works correctly with our library by running
python -m pytest tests/mujoco
Our library also contains a set of
visualization tools, meant to facilitate diagnostics and
development of models and controllers. These currently require a Mujoco
installation (see previous subsection), but we are planning to add support for other environments
and extensions in the future. Currently, the following tools are provided:
Visualizer: Creates a video to qualitatively
assess model predictions over a rolling horizon. Specifically, it runs a
user specified policy in a given environment, and at each time step, computes
the model's predicted observation/rewards over a lookahead horizon for the
same policy. The predictions are plotted as line plots, one for each
observation dimension (blue lines) and reward (red line), along with the
result of applying the same policy to the real environment (black lines).
The model's uncertainty is visualized by plotting lines the maximum and
minimum predictions at each time step. The model and policy are specified
by passing directories containing configuration files for each; they can
be trained independently. The following gif shows an example of 200 steps
of pre-trained MBPO policy on Inverted Pendulum environment.
DatasetEvaluator: Loads a pre-trained model
and a dataset (can be loaded from separate directories), and computes
predictions of the model for each output dimension. The evaluator then
creates a scatter plot for each dimension comparing the ground truth output
vs. the model's prediction. If the model is an ensemble, the plot shows the
mean prediction as well as the individual predictions of each ensemble member.
FineTuner: Can be used to train a
model on a dataset produced by a given agent/controller. The model and agent
can be loaded from separate directories, and the fine tuner will roll the
environment for some number of steps using actions obtained from the
controller. The final model and dataset will then be saved under directory
subdiris provided by the user.
True Dynamics Multi-CPU Controller: This script can run
a trajectory optimizer agent on the true environment using Python's
multiprocessing. Each environment runs in its own CPU, which can significantly
speed up costly sampling algorithm such as CEM. The controller will also save
a video if the
renderargument is passed. Below is an example on
HalfCheetah-v2 using CEM for trajectory optimization.
TrainingBrowser: This script launches a lightweight
training browser for plotting rewards obtained after training runs
(as long as the runs use our logger).
The browser allows aggregating multiple runs and displaying mean/std,
and also lets the user save the image to hard drive. The legend and axes labels
can be edited in the pane at the bottom left. Requires installing
Thanks to a3ahmad for the contribution.
Note that, except for the training browser, all the tools above require Mujoco
installation and are specific to models of type
We are planning to extend this in the future; if you have useful suggestions
don't hesitate to raise an issue or submit a pull request!
Please check out our documentation
and don't hesitate to raise issues or contribute if anything is unclear!