fast_irl

Contains PyTorch implementation of the FILTER algorithm for fast inverse reinforcement learning.

Running Experiments

To train an expert, run:

python experts/train.py -e env_name

To train a learner, run:

python learners/train.py -a algo_name -e env_name -s seed

This package supports training via:

• Behavioral Cloning (bc)
• Moment Matching (mm)
• FILTER(NR) (filter-nr)
• FILTER(BR) (filter-br)

on the following environments:

• HalfCheetahBulletEnv-v0 (halfcheetah)
• HopperBulletEnv-v0 (hopper)
• WalkerBulletEnv-v0 (walker)
• antmaze-large-play-v2 (antmaze).

For the first three environments, we use Soft-Actor Critic as our baseline policy optimizer. For antmaze, we use T3D+BC. See learners/gym_wrappers.py for wrappers to speed up learning for your own inverse reinforcement learning algorithms.