kristery / Imitation-Learning-from-Imperfect-Demonstration

[ICML 2019] Implementation of "Imitation Learning from Imperfect Demonstration"

Imitation Learning from Imperfect Demonstration

The TRPO part is hugely based on: https://github.com/ikostrikov/pytorch-trpo

Requirement

• Python 3.6
• PyTorch 0.4.1
• gym
• mujoco
• numpy
• scipy

Execute

The .py files take trajectories and confidence data as inputs (in demonstrations folder) and record accumulated reward at each update in the log folder. Please follow below commands to run our methods and baselines. Traj-size option is the same as specifying $n_c+n_u$ in the paper and num-epochs specifies the maximum number of update iterations.

• IC_GAIL

python IC_GAIL.py --env Ant-v2 --num-epochs 5000 --traj-size 600 

• 2IWIL

python 2IWIL.py --env Ant-v2 --num-epochs 5000 --traj-size 600 --weight

• GAIL (U+C)

python 2IWIL.py --env Ant-v2 --num-epochs 5000 --traj-size 600

• GAIL (C)

python 2IWIL.py --env Ant-v2 --num-epochs 5000 --traj-size 600 --weight --only --noconf

• GAIL (Reweight)

python 2IWIL.py --env Ant-v2 --num-epochs 5000 --traj-size 600 --weight --only