SEER

SEER - Combining Analytic Control with Learning to Create a Stabilizing Controller that Works in Reality

Built by Damjan Denic, Martin Graf, Nav Leelarathna, and Sepand Dyanatkar

Setup

1. Clone the repository inclusive submodules git clone --recurse-submodules [email protected]:Paralelopipet/SEER.git
2. Pretrained weights are available here.
   • If you want to use them, place them in the folder pybullet_multigoal_implementation/drl_implementation/examples
3. To setup natively:
   1. Create conda environment conda env create -f environment.yml
      • note: this will take a while (3 to 4 coffees)
   2. Activate environment conda activate l32_seer
   3. Install our gym package pip install --editable pybullet_multigoal_gym
   4. Install our RL package pip install --editable pybullet_multigoal_implementation
   5. Install seer package pip install --editable .
4. To setup with Docker:
   1. Run docker build -f evaluate.Dockerfile -t seer-evaluate .
      • this needs to be repeated whenever any files were changed

Test (requires native install)

Run pytest

Train

• to train natively, run python seer/train_and_eval_configs/config_runner.py --config seer.train_and_eval_configs.rl_training.<scenario to train>
• To train using Docker, run docker run --memory=6g --cpus=4 --mount "type=bind,source=$PWD/pybullet_multigoal_implementation/drl_implementation/examples,target=/root/pybullet_multigoal_implementation/drl_implementation/examples" -it seer-evaluate --config seer.train_and_eval_configs.rl_training.<scenario to train>
  • in CMD, replace $PWD by the absolute path of this directory
• In both cases, replace <scenario to train> with the name of the Python file containing the scenario you want to train (i.e., rl_config_train_basic)
• weights are saved to the pybullet_multigoal_implementation/drl_implementation/examples folder

Evaluate

• to evaluate natively, run python seer/train_and_eval_configs/config_runner.py --config seer.train_and_eval_configs.<scenario to evaluate>
• To train using Docker, run docker run --memory=4g --cpus=2 --mount "type=bind,source=$PWD/pybullet_multigoal_implementation/drl_implementation/examples,target=/root/pybullet_multigoal_implementation/drl_implementation/examples" -it seer-evaluate --config seer.train_and_eval_configs.<scenario to evaluate>
  • in CMD, replace $PWD by the absolute path of this directory
• In both cases, replace <scenario to evaluate> with the partial package name of the Python file containing the scenario you want to evaluate (i.e., rl_eval.basic.rl_config_eval_basic or baseline.baseline_config_noisy_slope)
• if evaluating the reinforcement learning solution, weights need to be present in the pybullet_multigoal_implementation/drl_implementation/examples folder

Debug in VSCode

• press F5 (launch configuration RL Trainer)
• or try out the other launch configurations in .vscode/launch.json