The traditional AI agents that we see in the news, such as AlphaGo, Pokerbot were trained in well-scoped environments, but as we look to the future, we want AI that are capable of working towards more complex tasks in richer environments. We’ve seen some really amazing environments and agents pop-up recently, so let’s explore them! Across these environments, agents have been programmed to acquire knowledge, have complex social interactions, learn how to interact with the physical world and been thrown in the deep-end with virtually unlimited access to the world’s resources.
In this article we’ll cover:
- Minecraft environments: acquiring knowledge
- Simulacra / Sims: social interactions
- Isaac Gym: physical reality
- The internet: AutoGPT’s playground
Minecraft: Acquiring knowledge
Minecraft is a popular sandbox video game that allows players to build and explore virtual worlds made up of blocks. In survival mode, players must gather resources, craft tools, and build structures to protect themselves from monsters that come out at night. They can also explore vast landscapes, mine for valuable resources, and engage in farming and crafting activities. All of this provides a rich environment for AI to explore.
There have been many popular environments built around Minecraft to further AI research. These include Malmo, MineRL and MineDojo. The latest of these, MineDojo hosts an impressive corpus of wikis, videos and tasks available to learn from.
The creators of this environment have also published Voyager, an LLM-powered embodied agent that explores the world, acquiring a diverse set of skills and making novel discoveries. The model operates with three modules:
- A curriculum to maximise discovery
- A skill library to store and retrieve complex behaviours from.
- An iterative prompting mechanism that generates executable code for embodied control.
Simulacra
Another core trait of our intelligence is the ability to interact with the people around us. In this paper, the authors have created LLM-powered agents and let them interact in an embodied state in a more realistic environment.
The environment mimics a small village of 25 inhabitants, each with a short backstory. The agents can interact with each other through natural language or with the environment by their actions. Within the village there are the common places we would expect to see: a cafe, bar, park, school and stores. Within these there are specific functional areas that agents can interact with such as a kitchen in a house and a stove in a kitchen.
The core features of the agents are:
- Memory stream: a collection of observations and thoughts that act as a history of the agent’s actions and thoughts.
- Reflection: a high-level abstract thought generated by the agent based on the memory stream.
- Planning: a high-level plan that the agent will act on, it is stored in the memory stream.
Isaac Gym: Physical Reality
Ultimately, intelligence we will want intelligence to be grounded in reality, because that is the world that more advanced AI will need to operate in. It’s great to be able to win at a human-invented game such as Starcraft, but it’s much more valuable for our society if AI can ‘win’ at the game of factory logistics. To that end, Isaac Gym is an environment that models physics of the world for the training of robotics tasks such as bipedal walking and hand control. It leverages PhysX, which is the same simulation engine used in many popular video games and Hollywood movies. The environment provides the ability to run multiple simulations concurrently, such as eight robotic hands operating in one environment, paving the way for more complex and diverse training scenarios.
The Internet: AutoGPT Unleashed
The above environments all have something in common - they’re restricted to a set of options well defined by the creators. The AutoGPT agent breaks this trend. With plug-ins, AutoGPT has access to the indescribable complexity that is the internet. These plug-ins range from specific ones such as Bing search, or email sending all the way to generic API tools which allow agents to interact with any service on the internet. This can not be understated, with API integrations, AI agents have the ability to not just read and receive all sorts of information (GET requests) but to influence the world (POST requests) including physical objects within it, such as your home lighting system.
The Next Generation of AI
The new wave of AI development is not just about improving the agents themselves, but about creating the environments in which they can learn, grow, and interact. From virtual worlds to physical simulations to the vast expanse of the internet, these environments are becoming increasingly complex and dynamic. This opens up a world of possibilities for what AI can do and how it can impact our lives, moving us closer towards more generally intelligent agents (AGI).
In conclusion, we are making remarkable strides in the field of AI. Our continued exploration and experimentation in diverse environments such as Minecraft, Simulacra and Isaac Gym, are broadening the boundaries of what’s possible. As we continue to innovate, we can anticipate a future where AI agents are capable of complex tasks and interactions, setting the stage for the next generation of AI.