mit-acl/rl_collision_avoidance

This is the training code for:

Journal Version: M. Everett, Y. Chen, and J. P. How, "Collision Avoidance in Pedestrian-Rich Environments with Deep Reinforcement Learning", in review, Link to Paper

Conference Version: M. Everett, Y. Chen, and J. P. How, "Motion Planning Among Dynamic, Decision-Making Agents with Deep Reinforcement Learning", IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018. Link to Paper, Link to Video

The gym environment code is included as a submodule.

Install

Grab the code from github, initialize submodules, install dependencies and src code

# Clone either through SSH or HTTPS (MIT-ACL users should use GitLab origin)
git clone --recursive [email protected]:mit-acl/rl_collision_avoidance.git

cd rl_collision_avoidance
./install.sh

There are some moderately large (10s of MB) checkpoint files containing network weights that are stored in this repo as Git LFS files. They should automatically be downloaded during the install script.

Train RL (starting with a network initialized through supervised learning on CADRL decisions)

To start a GA3C training run (it should get approx -0.05-0.05 rolling reward to start):

./train.sh TrainPhase1

To load that checkpoint and continue phase 2 of training, update the LOAD_FROM_WANDB_RUN_ID path in Config.py and do:

./train.sh TrainPhase2

By default, the RL checkpoints will be stored in RL_tmp and I think files will get overwritten if you train multiple runs. Instead, I like using wandb as a way of recording experiments/saving network parameters. To enable this, set the self.USE_WANDB flag to be True in Config.py, then checkpoints will be stored in RL/wandb/run-<datetime>-<id>.

To run experiments on AWS

Start a bunch (e.g., 5) of AWS instances -- I used c5.2xlarge because they have 8vCPUs and 16GB RAM (somewhat like my desktop?). Note: this is just an example, won't work out of box for you (has hard-coded paths)

Add the IP addresses into ga3c_cadrl_aws.sh.

./ga3c_cadrl_aws.sh panes
# C-a :setw synchronize-panes -- will let you enter the same command in each instance

Then you can follow the install & train instructions just like normal. When training, it will prompt you for a wandb login (can paste in the authorization code from app.wandb.ai/authorize).

Observed Issues

If on OSX, when running the ./train.sh script, you see:

objc[39391]: +[__NSCFConstantString initialize] may have been in progress in another thread when fork() was called.
objc[39391]: +[__NSCFConstantString initialize] may have been in progress in another thread when fork() was called. We cannot safely call it or ignore it in the fork() child process. Crashing instead. Set a breakpoint on objc_initializeAfterForkError to debug.

just add this ENV_VAR: export OBJC_DISABLE_INITIALIZE_FORK_SAFETY=YES.

If you find this code useful, please consider citing:

@inproceedings{Everett18_IROS,
  address = {Madrid, Spain},
  author = {Everett, Michael and Chen, Yu Fan and How, Jonathan P.},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  date-modified = {2018-10-03 06:18:08 -0400},
  month = sep,
  title = {Motion Planning Among Dynamic, Decision-Making Agents with Deep Reinforcement Learning},
  year = {2018},
  url = {https://arxiv.org/pdf/1805.01956.pdf},
  bdsk-url-1 = {https://arxiv.org/pdf/1805.01956.pdf}
}