PyTorch implementation of our work: "Sample-Efficient Imitation Learning via Generative Adversarial Nets". A TensorFlow implementation is also available at sam-tf.

Published in AISTATS 2019 | arXiv link | Video demos | Expert demonstrations

Launching scripts are available in /tasks. To run a task, use:

python spawner.py \
    --config tasks/sam_fixed_local_mujoco.yaml \
    --no-sweep \
    --call \
    --visdom_server=<visdom_server> \
    --visdom_port=<visdom_port> \
    --visdom_username=<visdom_username> \
    --visdom_password=<visdom_password>

Visdom can be disabled (and the associated options can therefore be omitted) by setting the argument enable_visdom to false in the YAML configuration file. The command triggers the creation of a tmux session in which jobs run in distinct windows. For example, if the configuration file specifies 3 environments and a number of random seeds equal to 4, the tmux session will have 12 windows, each containing its unique pair of environment and random seed.

• Interaction simulated with the openai/gym API.
• Some utilities were inspired from openai/baselines.