This repository contains code for training and evaluating goal-conditioned reinforcement learnin agents in BabyAI environments using different algorithms and configurations.
Clone the repository:
git clone https://github.com/your-username/rl-project.git
Create conda env
conda create -n myenv python=3.12
Activate conda env
conda activate myenv
Install the required dependencies:
pip install -r requirements.txt
To train an RL agent, you can run the main script with the desired command-line arguments. For example:
python testBabyaiEnv.py --env custom-dynamic --obs fully-observable --num-conv-layers 5 --num_env=32 --num-timesteps 4000000 --size 5 --reward-shaping True
This command will train an agent using the PPO algorithm on our custom 5x5 Goal Conditioned environment with a feature extractor of 5 conv layers. It will use the fully-observable observation and use reward shaping.
You can customize the training process by modifying the command-line arguments. Here are some of the available options:
--env: Environment type (e.g.,custom-set-goal,custom-dynamic,room)--obs: Observation type (e.g.,one-hot,img,fully-observable,fully-observable-one-hot)--algorithm: RL algorithm (e.g.,PPO,A2C,DQN,HER)--policy: Policy type (e.g.,CnnPolicy,MlpPolicy)--num-timesteps: Number of timesteps to train for--num-conv-layers: Number of convolutional layers in the policy (3, 5, or 8)--num_envs: Number of parallel environments to run--size: Size of the square environment--goal-features: Goal feature representation (e.g.,fully-observable,one-pos,same-network-fully-obs,HER)--reward-shaping: Enable reward shaping (default: False, don't include the flag if you don't want reward shaping)
Note that not all combinations of arguments are valid. For example using HER is only valid in the custom-dynamic environment. Invalid combinations may produce assertation errors or raise exceptions.
do make board env=target_env to create good logs of for the base target_env.
To start the associated tensorboard, run
tensorboard --logdir ./path/to/clean/dir