We are conducting an interesting experiment at Space Zero.
Translated by GPT
Space Zero is conducting an interesting experiment.
Recently, we have been developing an AI agent that roams around Space Zero. This agent thinks and acts autonomously. By using LLM and reinforcement learning, we are creating an agent indistinguishable from human players.
It recognizes surrounding objects and walks around through commands. We create terrains similar to reality for physical interactions. It’s an experiment with limitless possibilities.
In fact, expansion into physical AI is also possible. Given that Space Zero is an open world that generates all sorts of 3D models using AI, it’s suitable for simulating the complexity of the real world. If the AI model interacts with various automatically labeled 3D models to set and execute real goals, this signifies an expansion into physical AI. Learning in virtual space connects to learning in the real world.
I believe the era of LLM AI is almost concluded. It has become as essential as electricity. We are entering an era of LLM AI optimized for specific purposes. Following this, I anticipate the era of robotics will come.
Ultimately, we need AI that understands the real world. We aim to lay this groundwork first at Space Zero. When humanoid robots like Tesla Bot become widespread, we can imagine a time when robots interacting in virtual space are directly migrated.
Technical Separation -> Open Source Policy
For scalability, we developed in Python. After considering various libraries for physical simulation, we chose PyBullet. By specifying personas and behaviors with LLM and performing specific actions through reinforcement learning, we can achieve movements indistinguishable from humans.
Additionally, thanks to the recently implemented Rust game server, we resolved development bottlenecks. Since the server is separated, players can be launched anytime from other clients as long as they can authenticate. Overcoming technical barriers has made it easier to add external elements like agents.
We are integrating the strengths and weaknesses of each language into our product. While functioning independently, the project is separated, which is advantageous for future open-source strategies.