Awesome

Reproducing this in 2022

Here are some tips to follow to sample trajectories:

1. Recommendation: Make your own Conda environment
2. The correct Mujoco version is Mujoco200, do not try to use the latest one. https://www.chenshiyu.top/blog/2019/06/19/Tutorial-Installation-and-Configuration-of-MuJoCo-Gym-Baselines/ can you help you out here, get the free license here http://www.roboti.us/license.html Fair warning: the new mujoco versions have a different provider now(DeepMind).
3. Do NOT install dm_env via pip install dm_env, clone the repo and then run pip install wherever/you/saved/dm_env. These are two different packages. The same thing goes for dm_control.
4. Before you install either of these do install the requirements pip install -r requirements.txt, for all 3 packages respectively

Other hickups that got into my way:

• undefined symbol: __glewBindBuffer found the solution here https://github.com/openai/mujoco-py/issues/549, terminal solution sudo apt-get install libglfw3 libglew2.1 solved it for me
• protobuf: https://stackoverflow.com/questions/72441758/typeerror-descriptors-cannot-not-be-created-directly

It takes a lot of patience and frustration, but it's worth it.

Contrastive Forward Model

This is code to reproduce experiments for the paper Learning Predictive Representations for Deformable Objects Using Contrastive Estimation.

Installation

This project was run using Python 3.7.6. All the dependencies are in the requirements.txt file and we recommend creating a virtual environment and then installing by pip install -r requirements.txt.

You will also need to install a custom dm_env package and a custom dm_control package which has the relevant rope and cloth environments. You must use the cfm branch in the custom dm_control repo. Note that dm_control requires the Mujoco simulator to use. Finally, you will need to install this repo as a pip package: cd contrastive-forward-model; pip install -e .

Running

The steps to collect and run data are as follows. You may use the -h flag to show more customizable options.

1. Collect data by running python sample_trajectories.py
2. Process the data using python process_dataset.py data/rope
3. Train CFM with python run_train.py. You can customize your own flags to run it with different hyperparameters. The output is stored in the out/ folder
4. Run the evaluations with python run_evaluation.py out/*, which will generate json files and store them in out/<exp_name>/eval/<eval_name>/eval_results.json

Visualization

There are two ways to visualize your results. If you group up your result folders by seed, and store them in a single file, i.e. tmp, you may call python cfm/visualize/print_evaluation_stats.py tmp, which will print out the results in a formatted manner, with standard statistics across seeds.

If you are performing hyperparmeters tuning, it may be easier to run python cfm/visualize/to_csv.py out, which will generate progress.csv and params.json files in each eval directory. Then, you can use the rllab viskit library to view: python <path to viskit>/viskit/frontend.py out, where you can split by different hyperparameters and average over seeds.

Baselines

You can run the baselines by executing python run_baselines/run_train_<baseline_name>.py for step 3 instead of the CFM script. The rest of the steps are identical.