This implementation is based on the paper by: Lynch et al., "Learning Latent Plans from Play", 2019. https://learning-from-play.github.io/

The purpose of this project is to make a robot capable of performing many different complex tasks without having to relearn each from scratch in a continuous setting.

You can find our trained models and video results of our implementation in the following link https://drive.google.com/drive/folders/1X2LBbLa2RBjLQEEP1bqxm035ok_Oab34?usp=sharing

Our current model was trained over 1 day and 16 hours, you can visualize the plots inside analysis/plots or in tensorboard by running the following command

foo@bar:~$ tensorboard --logdir ./runs

To run this project it is required to have a working installation of Mujoco 2.0, further information can be found on the following webpage: https://www.roboti.us/index.html

• Train:
  • Pytorch: https://pytorch.org/get-started/locally/
  • Tensorboard: pip install tensorboard
  • Numpy: pip install numpy
  • Matplotlib: pip install matplotlib
  • Scipy: pip install scipy
  • CV2: pip install opencv-python
• Additional packages for running the test environment:
  • tqdm: pip install tqdm
  • mjrl: https://github.com/aravindr93/mjrl
  • puppet: https://github.com/vikashplus/puppet
  • relay-policy-learning env: https://github.com/google-research/relay-policy-learning

Please make sure that the "mjrl", "puppet" and "relay-policy-learning" folders are in the root folder of the project before executing the environment.

To train a model you will need to define the appropiate constant values in "/utils/constants.py" after that execute the command:

foo@bar:~$ python training.py

The model with the best accuracy and best training loss will be saved under the folder "/models".

We created this script to analyze the latent plan space representation, before calling this, make sure to change the reference to the appropiate path of the model.

foo@bar:~$ python tsne_analysis.py

The resulting plots and images will be saved under "analysis/temporal_tsne".

To visualize the model performance on a single task run the following script:

foo@bar:~$ python test_single_goal.py

To visualize the model performance on consecutive tasks run the following script, press key 'n' to switch the goal to the next image in the list and press 'e' if you like to end the episode.

foo@bar:~$ python test_consecutive.py

Before calling any test script, please make sure to change the reference to the appropiate path of the model. To further modify the specifications please check the python files which include a more detailed information on the parameters used. Videos will be saved under "./analysis/videos/model_trials/goal_name".

For any other question or feedback, please send an email to:

• erickrosetebeas@hotmail.com
• jessicalizethborja@gmail.com
• guttikondasuresh7@gmail.com

We will be happy to help you.