DRL code based on https://github.com/openai/universe-starter-agent

It is for the training code of the problem of active tracking discussed in the following paper.

@inproceedings{luo2018end,
title={End-to-end Active Object Tracking via Reinforcement Learning},
author={Luo, Wenhan and Sun, Peng and Zhong, Fangwei and Liu, Wei and Zhang, Tong and Wang, Yizhou},
booktitle={International Conference on Machine Learning},
year={2018}
}

Dependencies

See https://github.com/openai/universe-starter-agent

The followings are optional

• gym_tvizdoom (for vizdoom related envs)
• gym_unrealcv (for object tracking based on unreal engine envs)

How to run

vizdoom example

python train.py --num-workers=2 --env-id=TrackObjSmallMazeRandFlip-v2 --log-dir=save/to --max-global-steps=29999 --val-model-secs=30 --max-val-episodes=2

python evaluate.py --env-id=TrackObjCounterclockwise-v2 --ckpt-dir=save_gallery/tosmrandflip-t8-v2-val --sleep-time=0.00 --max-episodes=10 --render --verbose=1

UnrealCV example

python train.py --num-workers=1 --env-id=Active-Tracking-Discrete-v0 --env-val-id=Active-Tracking-Discrete-v0 --model-id=convx2lstm --lr=0.00005 --log-dir=save/tmp --max-global-steps=250000000 --val-model-secs=-1 --max-val-episodes=40

python train.py --num-workers=1 --env-id=Tracking-Indoor1JasperPath1Static-v0 --env-val-id=Tracking-Indoor1JasperPath1Static-v0 --model-id=convx2lstm_small --lr=0.00005 --max-global-steps=250000000 --log-dir=save/tmp --val-model-secs=-1 --max-val-episodes=40