This github: https://github.com/mvieth/inverse_rl
The original authors' github: https://github.com/justinjfu/inverse_rl
Implementations for imitation learning / IRL algorithms in RLLAB
Contains:
- GAIL (https://arxiv.org/abs/1606.03476/pdf)
- Guided Cost Learning, GAN formulation (https://arxiv.org/pdf/1611.03852.pdf)
- Tabular MaxCausalEnt IRL (http://www.cs.cmu.edu/~bziebart/publications/thesis-bziebart.pdf)
This library requires:
- rllab (https://github.com/openai/rllab)
- Tensorflow
Running the Pendulum-v0 gym environment:
- Collect expert data
python scripts/pendulum_data_collect.py
You should get an "AverageReturn" of around -100 to -150
- Run imitation learning
python scripts/pendulum_irl.py
The "OriginalTaskAverageReturn" should reach around -100 to -150
- Run GAIL as a comparison
python scripts/pendulum_gail.py
- Collect expert data
python scripts/ant_data_collect.py
- Run AIRL
python scripts/ant_irl.py
- Transfer to a disabled ant using the parameters from step 2
python scripts/ant_transfer_disabled.py
- Load a policy, execute it, and capture the video
python scripts/ant_demonstrate.py
The data
folder contains data from the executed scripts. The pickle files (.pkl) contain the learned policies, among other things.