Wongcheukwai/taco

TACO: Learning Task Decomposition via Temporal Alignment for Control

TACO is a learning from demonstration (LfD) algorithm that can be used to learn simple sub policies from complex demonstrations augmented by a task sketch. Paper: https://arxiv.org/pdf/1803.01840.pdf

Dependencies

Tensorflow, MujocoPy Numpy, Scipy, Pandas, Seaborn, Matplotlib.

Datasets

A small dataset for NavWorld can be found in data/nav/dataset_04.p.

You can also collect your own dataset from these domains.

For NavWorld:

	python3 nav_world.py [dataset_name] -n 1000

For Dial:

	python3 jacopinpad_collect collect [dataset_dir] -n [number of datapoints] -l [sketch_length] --permute

For Dial (visual):

	python3 jacopinpad_collect collect [dataset_dir] -n [number of datapoints] -l [sketch_length] --img_collect

Data should go into the respective domain folder in the data folder.

Reproducing:

Add the taco directory to your ~/.bashrc. To run using the example dataset use

python3 experiment_launch.py reproduce_taco nav taco dataset_04.p -c taco_nav_base.yaml -n 400

To run using multiple methods use the .sh scripts for the nav domain.

Evaluating:

After training is finished:

python3 evaluation.py ../results/reproduce_taco nav 

This will also plot the results for all the models trained within the test_nav domain. Use the -r flag to render the evaluation to a video (for nav only).