Code for the paper: Motion Perception in Reinforcement Learning with Dynamic Objects
Artemij Amiranashvili, Alexey Dosovitskiy, Vladlen Koltun and Thomas Brox, CoRL 2018 (paper link).
Dependencies:
- Python3
- TensorFlow
- mujoco-py
Installation:
Run pip3 install -e . in the gym and baselines directories.
Tasks:
- Chaser2d-v2
- Catcher2d-v2
- Catcher3d-v1
- KeepUp3d-v1
- ChaserWithRandom4Backgrounds2d-v2
- KeepUpHighMotionPenalty3d-v1
Running Examples:
Training Catcher3d-v1 with additional TinyFlowNet flow prediction input (replace LOG_DIR with path for logging):
python3 baselines/baselines/ppo2/run_mujoco_imvec.py --main_path LOG_DIR --env_id Catcher3d-v1 --add_flownet True --flownet_path networks/Catcher3d/
Training Chaser2d-v2 with image stack input:
python3 baselines/baselines/ppo2/run_mujoco_imvec.py --main_path LOG_DIR --stack_frames True
Training Chaser2d-v2 with additional image difference input:
python3 baselines/baselines/ppo2/run_mujoco_imvec.py --main_path LOG_DIR --stack_frames True --diff_frames True