Inverse Dynamics Simulation for Learning Experiments

This repository contains a very simple but easy to extend simulation which enables us to run various inverse dynamics learning experiments, and compare different approaches.

All following commands are assumed to be executed from the idsim root directory.

Installation

For ease of use an installation script is provided which sets up a virtual environement. We install all necessary dependencies into this virtual environment.

The user has the option to either use tensorflow CPU or GPU. Note, one of the two versions has to be selected.

bash scripts/install.sh

Iros Experiments

In order to run the Iros experiments please execute the following script:

bash scripts/exp_iros/create_experiment.sh

This will create a folder structure in logs.

ls logs/exp_iros/

Should now show a lot of folders with descriptive names for every single experiment. Every single experiment folder should contain the following files

config.yaml       # All open parameters used for this experiment.
experiment.sh     # The script to run exactly this experiment.
results.sh        # The script to generate the data required to aggregate the results.
template_params.yaml  # The template parameters used to generate the config.yaml.
visualize.sh      # This script will visualize how the system evolves over time, and a lot of other properties ofr this experiment.

Most notably there are 4 files prefixed with cluster_. The two files called cluster_exp.sub and cluster_res.sub can be used in a htcondor cluster environment. This will run all experiments in parallel.

However, experiments can be run manually and sequential by using the scripts cluster_exp.sh and cluster_res.sh.

# The script cluster_exp.sh takes a single parameter, namely the index of the experiment.
# In this particular cases there exist 1280 experiments, thus, every single one can be run
# as follows.
# experiment 0;
bash logs/exp_iros/cluster_exp.sh 0 
# experiment 1;
bash logs/exp_iros/cluster_exp.sh 1

# A simple sequential execution could look like this;
for i in $(seq 1 1280);do
  bash logs/exp_iros/cluster_exp.sh $i
  bash logs/exp_iros/cluster_res.sh $i
done

In order to create the plots from the paper we have to aggregate all results.

source vpy/bin/activate
python inverse_dynamics/results.py \
  --dp_exp_root logs/exp_iros/ \
  --dp_output logs/exp_iros/results \
  --plot \
  --average_params '[NOISE_FORWARD_FRICTION, NOISE_FORWARD_STICTION, NOISE_SENSING, RANDOM_SEED]' \
  --q \
  --tau_feedback_applied  

The resulting plots will be located in logs/exp_iros/results.