VectorVelocity is a space-themed OpenAI Gym environment designed for reinforcement learning. Users pilot a spaceship through an asteroid field while collecting coins. The environment increases in difficulty as the speed escalates, with asteroids and coins spawning randomly, providing dynamic challenges that are ideal for developing and refining Agents
In VectorVelocity, the player controls a spaceship moving across three lanes. The objective is to collect as many coins as possible while avoiding collisions with asteroids that move from the top of the screen to the bottom. As the game progresses, the speed increases, making the game increasingly difficult.
This game was enriched significantly by incorporating various third-party assets. We are immensely grateful to the creators of these assets for making their work available and enhancing the gaming experience.
We extend a huge thanks to the mentioned authors for making their work freely available.
To install the Vector Velocity Environment, you can use pip. Simply run the following command in your terminal:
pip install vector-velocity-gym
After installing the environment, you may want to test the installation and explore how to build or use the environment effectively.
For comprehensive guides and examples, please visit our offical documentation
The challenge for the RL agent in VectorVelocity is to learn optimal strategies for maximizing the score by skillfully collecting coins while avoiding asteroids. The agent is required to make decisions in real-time, adjusting to the game's increasing speed and the randomness of asteroid placements. Additionally, some coins spawn between asteroids in positions that may not always be reachable, adding a layer of decision-making complexity. This requires the player, and consequently the RL agent, to assess whether pursuing a coin is worth the risk of potential collision. This problem domain provides a rich and challenging environment for exploring and refining reinforcement learning techniques.
We have decided to use the Proximal Policy Optimization (PPO) algorithm to train an agent within our environment. PPO, a reinforcement learning algorithm developed by OpenAI, is known for its robustness and efficiency in learning policies for various types of environments.
For those interested in seeing a practical implementation or experimenting with the agent development process, sample agent development can be found in the lab
branch of this repository. This branch includes experimental features and developmental progress on new agent strategies.
To access and contribute to the ongoing agent development, switch to the lab
branch:
We welcome everyone to contribute to the project. Share your ideas, engage in discussions, and help us improve VectorVelocity. To get started, please check out the CONTRIBUTING.md file in the repo for detailed guidelines.