RL-botics is a toolbox with highly optimized implementations of Deep Reinforcement Learning algorithms for robotics developed with Keras and TensorFlow in Python3.
The objective was to have modular, clean and easy to read codebase so that the research community may build on top with ease. The implementations can be integrated with OpenAI Gym environments. The majority of the algorithms are Policy Search Methods as the toolbox is targetted for robotic applications.
Requirements:
It is highly recommended to install this package in a virtual environment, such as Miniconda. Please find the Conda installation here.
To create a new conda environment called RL
:
conda create -n RL python=3
To activate the environment:
source activate RL
To deactivate the environment:
source deactivate
To install the package, we recommend cloning the original package:
git clone https://github.com/Suman7495/rl-botics.git
cd rl-botics
pip install -e .
To run any algorithm in the default setting, simply run:
cd rl_botics/<algo>/
python main.py
For example, to run TRPO:
cd rl_botics/trpo/
python main.py
Numerous other options can be added too, but it is recommended to modify the hyerperparameters in hyperparameters.py
.
The algorithms implemented are:
- Q-Learning
- Deep Q-Network
- Vanilla Policy Gradient
- Deep Deterministic Policy Gradient
- Trust Region Policy Optimization
- Proximal Policy Optimization
- Proximal Policy Optmization with Intrinsic Curiosity Module (ICM)
- Compatible Natural Policy Gradient
To be added:
All environments are in the envs
directory. The environments available currently are:
- Field Vision Rock Sampling (FVRS): A POMDP environment where the agent has to collect good rocks from partial observability.
- Table Continuous: A POMDP environment emulation Human Robot Collaboration. The objective of the robot is to remove dirty dishes from the table without colliding with the human.
All the algorithms are in the rl_botics
directory. Each algorithm specified above has an individual directory.
The directory common
contains common modular classes to easily build new algorithms.
approximators
: Basic Deep Neural Networks (Dense, Conv, LSTM).data_collection
: Performs rollouts and collect observations and rewardslogger
: Log training data and other informationplotter
: Plot graphspolicies
: Common policies such as Random, Softmax, Parametrized Softmax and Gaussian Policyutils
: Functions to compute the expected return, the Generalized Advantage Estimation (GAE), etc.
Each algorithm directory contains at least 3 files:
main.py
: Main script to run the algorithmhyperparameters.py
: File to contain the default hyperparameters<algo>.py
: Implementation of the algorithmutils.py
: (Optional) File containing some utility functions
Some algorithm directories may have additional files specific to the algorithm.
To contribute to this package, it is recommended to follow this structure:
- The new algorithm directory should at least contain the 3 files mentioned above.
main.py
should contain at least the following functions:main
: Parses input argument, builds the environment and agent, and train the agent.argparse
: Parses input argument and loads default hyperparameters fromhyperparameter.py
.
<algo>.py
should contain at least the following methods:__init__
: Initializes the classes_build_graph
: Calls the following methods to build the TensorFlow graph:_init_placeholders
: Initialize TensorFlow placeholders_build_policy
: Build policy TensorFlow graph_build_value_function
: Build value function TensorFlow graph_loss
: Build policy loss function TensorFlwo graph
train
: Main training loop called bymain.py
update_policy
: Update the policyupdate_value
: Update the value functionprint_results
: Print the training resultsprocess_paths
: (optional) Process collected trajectories to return the feed dictionary for TensorFlow
It is recommended to check the structure of ppo.py
and follow a similar structure.
Suman Pal
MIT License.