This project uses a DRL algorithm (TD3) to perform a goal-reaching task for a robotic arm, while avoiding obstacles using camera images. This is the result of my master thesis, which can be found here.

• Unity
• ML-Agents
• pytorch

To get a local copy up and running follow these simple steps.

This project was tested using python versions 3.7.10 and 3.8.8, but any recent python version should work.

1. Clone the repo

   git clone https://github.com/fpadula/visualcollisionarm

2. Navigate to the Python Scripts folder and create a new virtualenv

   cd python_scripts/
   python -m virtualenv venv

3. Enable the virtualenv and install all the necessary python packages

   source venv/bin/activate
   pip install -r python_packages.txt

1. Run a pre-trained model with a single agent using visual and scalar inputs, and enable input visualization

   python src/trainer.py --run_id VisualModel --config_file configs/base_config_visual_aug.yaml --device cuda --exec_type eval --env_location simu_envs/SingleAgentVisualization/scene.x86_64 --simu_spd 1 --eval_episodes 10 --seed 1 --visualize_input true

2. Run a pre-trained model using multiple agents that uses visual and scalar values

   python src/trainer.py --run_id VisualModel --config_file configs/base_config_visual_aug.yaml --device cuda --exec_type eval --env_location simu_envs/AllAgentsVisual/scene.x86_64 --simu_spd 1

3. Run a pre-trained model using multiple agents that uses only scalar values

   python src/trainer.py --run_id ScalarModel --config_file configs/base_config.yaml --device cuda --exec_type eval --env_location simu_envs/AllAgentsScalar/scene.x86_64 --simu_spd 1

4. Gym-Wrapper example; running a random policy:

   python examples/gym_api.py

• Organize the code a little bit better
• Add more usage examples