README.md

State Representation Learning Zoo with PyTorch (part of S-RL Toolbox)

A collection of State Representation Learning (SRL) methods for Reinforcement Learning, written using PyTorch.

SRL Zoo Documentation: https://srl-zoo.readthedocs.io/

S-RL Toolbox Documentation: https://s-rl-toolbox.readthedocs.io/

S-RL Toolbox Repository: https://github.com/araffin/robotics-rl-srl

Available methods:

• Autoencoder (reconstruction loss)
• Denoising Autoencoder (DAE)
• Forward Dynamics model
• Inverse Dynamics model
• Reward prediction loss
• Variational Autoencoder (VAE) and beta-VAE
• SRL with Robotic Priors + extensions (stereovision, additional priors)
• Supervised Learning
• Principal Component Analysis (PCA)
• Triplet Network (for stereovision only)
• Combination and stacking of methods
• Random Features
• [experimental] Reward Prior, Episode-prior, Perceptual Similarity loss (DARLA), Mutual Information loss

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
• "State Representation Learning for Control: An Overview" (Lesort et al., 2018), link: https://arxiv.org/pdf/1802.04181.pdf

Documentation

Documentation is available online: https://srl-zoo.readthedocs.io/

Installation

Please read the documentation for more details, we provide anaconda env files and docker images.

Learning a State Representation

To learn a state representation, you need to enforce constrains on the representation using one or more losses. For example, to train an autoencoder, you need to use a reconstruction loss. Most losses are not exclusive, that means you can combine them.

All losses are defined in losses/losses.py. The available losses are:

• autoencoder: reconstruction loss, using current and next observation
• denoising autoencoder (dae): same as for the auto-encoder, except that the model reconstruct inputs from noisy observations containing a random zero-pixel mask
• vae: (beta)-VAE loss (reconstruction + kullback leiber divergence loss)
• inverse: predict the action given current and next state
• forward: predict the next state given current state and taken action
• reward: predict the reward (positive or not) given current and next state
• priors: robotic priors losses (see "Learning State Representations with Robotic Priors")
• triplet: triplet loss for multi-cam setting (see Multiple Cameras section in the doc)

[Experimental]

• reward-prior: Maximises the correlation between states and rewards (does not make sense for sparse reward)
• episode-prior: Learn an episode-agnostic state space, thanks to a discriminator distinguishing states from same/different episodes
• perceptual similarity loss (for VAE): Instead of the reconstruction loss in the beta-VAE loss, it uses the distance between the reconstructed input and real input in the embedding of a pre-trained DAE.
• mutual information loss: Maximises the mutual information between states and rewards

All possible arguments can be display using python train.py --help. You can limit the training set size (--training-set-size argument), change the minibatch size (-bs), number of epochs (--epochs), ...

Datasets: Simulated Environments and Real Robots

Although the data can be generated easily using the RL repo in simulation (cf Generating Data), we provide datasets with a real baxter:

• Dataset 1
• Dataset 2 with multiple cameras

Examples

You can download an example dataset here.

Train an inverse model:

python train.py --data-folder data/path/to/dataset --losses inverse

Train an autoencoder:

python train.py --data-folder data/path/to/dataset --losses autoencoder

Combining an autoencoder with an inverse model is as easy as:

python train.py --data-folder data/path/to/dataset --losses autoencoder inverse

You can as well specify the weight of each loss:

python train.py --data-folder data/path/to/dataset --losses autoencoder:1 inverse:10

Please read the documentation for more examples.

Running Tests

Download the test datasets kuka_gym_test and kuka_gym_dual_test and put it in data/ folder.

./run_tests.sh

Troubleshooting

CUDA out of memory error

1. python train.py --data-folder data/staticButtonSimplest

RuntimeError: cuda runtime error (2) : out of memory at /b/wheel/pytorch-src/torch/lib/THC/generic/THCStorage.cu:66

SOLUTION 1: Decrease the batch size, e.g. 32-64 in GPUs with little memory.

SOLUTION 2 Use simple 2-layers neural network model python train.py --data-folder data/staticButtonSimplest --model-type mlp