Skip to content

S-RL Toolbox: Reinforcement Learning (RL) and State Representation Learning (SRL) for Robotics

License

Notifications You must be signed in to change notification settings

fengyuan1993012/robotics-rl-srl

 
 

Repository files navigation

S-RL Toolbox: Reinforcement Learning (RL) and State Representation Learning (SRL) Toolbox for Robotics

This repository was made to evaluate State Representation Learning methods using Reinforcement Learning. It integrates (automatic logging, plotting, saving, loading of trained agent) various RL algorithms (PPO, A2C, ARS, ACKTR, DDPG, DQN, ACER, CMA-ES, SAC, TRPO) along with different SRL methods (see SRL Repo) in an efficient way (1 Million steps in 1 Hour with 8-core cpu and 1 Titan X GPU).

We also release customizable Gym environments for working with simulation (Kuka arm, Mobile Robot in PyBullet, running at 250 FPS on a 8-core machine) and real robots (Baxter Robot, Robobo with ROS).

Related papers:

  • "Decoupling feature extraction from policy learning: assessing benefits of state representation learning in goal based robotics" (Raffin et al. 2018) https://arxiv.org/abs/1901.08651
  • "S-RL Toolbox: Environments, Datasets and Evaluation Metrics for State Representation Learning" (Raffin et al., 2018) https://arxiv.org/abs/1809.09369

Main Features

  • 10 RL algorithms (Stable Baselines included)
  • logging / plotting / visdom integration / replay trained agent
  • hyperparameter search (hyperband, hyperopt)
  • integration with State Representation Learning (SRL) methods (for feature extraction)
  • visualisation tools (explore latent space, display action proba, live plot in the state space, ...)
  • robotics environments to compare SRL methods
  • easy install using anaconda env or Docker images (CPU/GPU)

Documentation

Documentation is available online: https://s-rl-toolbox.readthedocs.io/

Example

Here is a quick example of how to train a PPO2 agent on MobileRobotGymEnv-v0 environment for 10 000 steps using 4 parallel processes:

python -m rl_baselines.train --algo ppo2 --no-vis --num-cpu 4 --num-timesteps 10000 --env MobileRobotGymEnv-v0

The complete command (logs will be saved in logs/ folder):

python -m rl_baselines.train --algo rl_algo --env env1 --log-dir logs/ --srl-model raw_pixels --num-timesteps 10000 --no-vis

To use the robot's position as input instead of pixels, just pass --srl-model ground_truth instead of --srl-model raw_pixels

Installation

Python 3 is required (python 2 is not supported because of OpenAI baselines)

Note: we are using Stable Baselines, a fork of OpenAI Baselines with unified interface and other improvements (e.g. tensorboard support).

Using Anaconda

  1. Download the project (note the --recursive argument because we are using git submodules):
git clone git@github.com:araffin/robotics-rl-srl.git --recursive
  1. Install the swig library:
sudo apt-get install swig
  1. Install the dependencies using environment.yml file (for anaconda users) in the current environment
conda env create --file environment.yml
source activate py35

PyBullet Documentation

Using Docker

Please read the documentation for more details.

Reinforcement Learning

Several algorithms from Stable Baselines have been integrated along with some evolution strategies and SAC:

  • A2C: A synchronous, deterministic variant of Asynchronous Advantage Actor Critic (A3C).
  • ACER: Sample Efficient Actor-Critic with Experience Replay
  • ACKTR: Actor Critic using Kronecker-Factored Trust Region
  • ARS: Augmented Random Search (https://arxiv.org/abs/1803.07055)
  • CMA-ES: Covariance Matrix Adaptation Evolution Strategy
  • DDPG: Deep Deterministic Policy Gradients
  • DeepQ: DQN and variants (Double, Dueling, prioritized experience replay)
  • PPO1: Proximal Policy Optimization (MPI Implementation)
  • PPO2: Proximal Policy Optimization (GPU Implementation)
  • SAC: Soft Actor Critic
  • TRPO: Trust Region Policy Optimization (MPI Implementation)

Please read the documentation for more details on how to train/load an agent on discrete/continuous actions, and how to add your own rl algorithm.

Hyperparameter Search

This repository also allows hyperparameter search, using hyperband or hyperopt for the implemented RL algorithms

for example, here is the command for a hyperband search on PPO2, ground truth on the mobile robot environment:

python -m rl_baselines.hyperparam_search --optimizer hyperband --algo ppo2 --env MobileRobotGymEnv-v0 --srl-model ground_truth

Environments

All the environments we propose follow the OpenAI Gym interface. We also extended this interface (adding extra methods) to work with SRL methods (see State Representation Learning Models).

Available Environments

Kuka environment Mobile Robot environment Racing car environment Omnidirectional robot environment
Name Action space (discrete) Action space (continuous) Rewards ground truth
Kuka
Button
6 actions (3D cardinal direction) 3 axis (3D cardinal direction) (1) 1 when target reached, -1 when too far from target or when table is hit, otherwise 0 (2) (3) the X,Y,Z position of the effector (4)
Kuka
RandButton
6 actions (3D cardinal direction) 3 axis (3D cardinal direction) (1) 1 when target reached, -1 when too far from target or when table is hit, otherwise 0 (2) (3) the X,Y,Z position of the effector (4)
Kuka
2Button
6 actions (3D cardinal direction) 3 axis (3D cardinal direction) (1) 1 when the first target is reached, 1 when the second target is reached, -1 when too far from target or when table is hit, otherwise 0 (2) the X,Y,Z position of the effector (4)
Kuka
MovingButton
6 actions (3D cardinal direction) 3 axis (3D cardinal direction) (1) 1 when target reached, -1 when too far from target or when table is hit, otherwise 0 (2) (3) the X,Y,Z position of the effector (4)
MobileRobot
4 actions (2D cardinal direction) 2 axis (2D cardinal direction) 1 when target reached, -1 for a wall hit, otherwise 0 (2) the X,Y position of the robot (4)
MobileRobot
2Target
4 actions (2D cardinal direction) 2 axis (2D cardinal direction) 1 when target reached, -1 for a wall hit, otherwise 0 (2) the X,Y position of the robot (4)
MobileRobot
1D
2 actions (1D cardinal direction) 1 axis (1D cardinal direction) 1 when target reached, -1 for a wall hit, otherwise 0 (2) the X position of the robot (4)
MobileRobot
LineTarget
4 actions (2D cardinal direction) 2 axis (2D cardinal direction) 1 when target reached, -1 for a wall hit, otherwise 0 (2) the X,Y position of the robot (4)
CarRacing 4 actions (left, right, accelerate, brake) 3 axis (stearing, accelerate, brake) -100 when out of bounds, otherwise -0.1 the X,Y position of the car (4)
OmniRobot 4 actions (2D cardinal direction) 2 axis (2D cardinal direction) 1 when target reached, -1 for a wall hit, otherwise 0 (2) the X,Y position of the robot (4)

1. The action space can use 6 axis arm joints control with the --joints flag
2. The reward can be the euclidian distance to the target with the --shape-reward flag
3. When using --shape-reward and --continuous, the reward for hitting the button is 50 and for being out of bounds is -250. This is to prevent the agent hitting the table to stop the environment early and obtaining a higher reward
4. The ground truth can be relative position from agent to the target by changing the RELATIVE_POS constant in the environment file

the available environments are:

  • Kuka arm: Here we have a Kuka arm which must reach a target, here a button.
    • KukaButtonGymEnv-v0: Kuka arm with a single button in front.
    • KukaRandButtonGymEnv-v0: Kuka arm with a single button in front, and some randomly positioned objects
    • Kuka2ButtonGymEnv-v0: Kuka arm with 2 buttons next to each others, they must be pressed in the correct order (lighter button, then darker button).
    • KukaMovingButtonGymEnv-v0: Kuka arm with a single button in front, slowly moving left to right.
  • Mobile robot: Here we have a mobile robot which reach a target position
    • MobileRobotGymEnv-v0: A mobile robot on a 2d terrain where it needs to reach a target position.
    • MobileRobot2TargetGymEnv-v0: A mobile robot on a 2d terrain where it needs to reach two target positions, in the correct order (lighter target, then darker target).
    • MobileRobot1DGymEnv-v0: A mobile robot on a 1d slider where it can only go up and down, it must reach a target position.
    • MobileRobotLineTargetGymEnv-v0: A mobile robot on a 2d terrain where it needs to reach a colored band going across the terrain.
  • Racing car: Here we have the interface for the Gym racing car environment. It must complete a racing course in the least time possible (only available in a terminal with X running)
    • CarRacingGymEnv-v0: A racing car on a racing course, it must complete the racing course in the least time possible.
  • Baxter: A baxter robot that must reach a target, with its arms. (see Working With Real Robots: Baxter and Robobo)
    • Baxter-v0: A bridge to use a baxter robot with ROS (in simulation, it uses Gazebo)
  • Robobo: A Robobo robot that must reach a target position.
    • RoboboGymEnv-v0: A bridge to use a Robobo robot with ROS.
  • OmniRobot: An Omnidirectional robot on a 2d terrain that must reach a target position (see Working With Real Robots: OmniRobot)
    • OmnirobotEnv-v0: Simulator but also a bridge to use an OmniRobot with ROS.

Please read the documentation for more details (e.g. adding a custom environment).

State Representation Learning Models

Please look the SRL Repo to learn how to train a state representation model. Then you must edit config/srl_models.yaml and set the right path to use the learned state representations.

The available state representation models are:

  • ground_truth: Hand engineered features (e.g., robot position + target position for mobile robot env)
  • raw_pixels: Learning a policy in an end-to-end manner, directly from pixels to actions.
  • autoencoder: an autoencoder from the raw pixels
  • inverse: an inverse dynamics model
  • forward: a forward dynamics model
  • vae: a variational autoencoder from the raw pixels
  • random: random features, the feature extractor, a convolutional network, is fixed after random initialization.
  • srl_combination: a model combining several losses (e.g. vae + forward + inverse...) for SRL
  • supervised: A model trained with Ground Truth states as targets in a supervised setting.
  • robotic_priors: Robotic Priors model
  • pca: pca applied to the raw pixels
  • multi_view_srl: a SRL model using views from multiple cameras as input, with any of the above losses (e.g triplet and others)
  • joints: the arm's joints angles (only for Kuka environments)
  • joints_position: the arm's x,y,z position and joints angles (only for Kuka environments)

Please read the documentation for more details (e.g. adding a custom SRL model).

Troubleshooting

If a submodule is not downloaded:

git submodule update --init

If you have troubles installing mpi4py, make sure you the following installed:

sudo apt-get install libopenmpi-dev openmpi-bin openmpi-doc

Known issues

The inverse kinematics function has trouble finding a solution when the arm is fully straight and the arm must bend to reach the requested point.

Acknowledgements

This work is supported by the DREAM project through the European Union Horizon 2020 FET research and innovation program under grant agreement No 640891.

Citation

If you use this toolbox, please cite:

@article{Raffin18,
  title={S-RL Toolbox: Environments, Datasets and Evaluation Metrics for State Representation Learning},
  author={Raffin, Antonin and Hill, Ashley and Traor{\'e}, Ren{\'e} and Lesort, Timoth{\'e}e and D{\'\i}az-Rodr{\'\i}guez, Natalia and Filliat, David},
  journal={arXiv preprint arXiv:1809.09369},
  year={2018}
}

About

S-RL Toolbox: Reinforcement Learning (RL) and State Representation Learning (SRL) for Robotics

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Python 99.8%
  • Shell 0.2%