TLDR: ONELIFE is a new AI framework that learns symbolic world models in complex, stochastic environments with limited, unguided exploration. It uses conditionally-activated programmatic laws with a precondition-effect structure to model dynamics. Evaluated on Crafter-OO, it outperforms baselines in predicting plausible future states and supports effective planning, laying groundwork for autonomous world model construction.
A new research paper introduces ONELIFE, a groundbreaking framework designed to teach artificial intelligence agents how to understand and predict the dynamics of complex, unpredictable virtual worlds. Unlike previous approaches that often rely on extensive data and human guidance, ONELIFE tackles the challenging scenario where an agent has only a single, unguided exploration “life” in a hostile, stochastic environment.
Understanding Complex Worlds with Limited Experience
The core idea behind ONELIFE is to infer and represent the transitional dynamics of an environment as an executable program, known as a symbolic world model. Imagine an AI agent dropped into a game like Minecraft or RuneScape, but without any instructions, rewards, or prior knowledge. ONELIFE aims to enable such an agent to reverse-engineer the game’s rules from minimal interaction.
The framework models world dynamics through what it calls “conditionally-activated programmatic laws” within a probabilistic programming setup. Each law works with a “precondition-effect” structure. This means a law only activates if certain conditions are met, and then it predicts only the specific attributes it directly governs. This smart design helps ONELIFE avoid the common scaling problems in complex environments, allowing it to learn stochastic dynamics accurately even when most rules are inactive at any given moment.
Crafter-OO: A New Testbed for AI Learning
To rigorously evaluate ONELIFE under these demanding constraints, the researchers developed Crafter-OO. This is a re-engineered version of the popular Crafter environment, specifically designed to expose a structured, object-oriented symbolic state and a pure transition function. This makes it an ideal testbed for symbolic world modeling, allowing AI models to “read” and “write” code that modifies the game state directly.
Measuring Understanding: Ranking and Fidelity
The evaluation of ONELIFE involved a new protocol with two key metrics: state ranking and state fidelity. State ranking measures the agent’s ability to distinguish between plausible and implausible future states. For example, if an agent tries to craft an item without the necessary resources, a good world model should recognize that the outcome of successfully crafting the item is “implausible.” State fidelity, on the other hand, assesses how closely the generated future states resemble reality.
In experiments conducted on Crafter-OO, ONELIFE successfully learned key environment dynamics from minimal, unguided interaction. It outperformed a strong baseline, PoE-World, in 16 out of 23 scenarios tested. This demonstrates its superior ability to predict true environment dynamics and distinguish valid outcomes from invalid ones.
Also Read:
- Ctrl-World: Advancing Robot Learning Through Imagined Interactions
- How2: A Framework for Lifelong Learning in AI Agents Through Procedural Questions
Planning for the Future
Beyond just understanding the world, ONELIFE also proved useful for planning. By simulating different strategies within its learned world model, the agent could successfully identify superior approaches for goal-oriented tasks. For instance, in a “Zombie Fighter” scenario, the model correctly predicted that crafting a sword before engaging in combat was a more effective strategy than fighting immediately. This highlights ONELIFE’s potential to support intelligent decision-making in unknown and dangerous environments.
This research lays a crucial foundation for autonomously constructing programmatic world models of unknown, complex environments, paving the way for more adaptable and intelligent AI agents. You can read the full paper here: ONELIFE TOLEARN: INFERRINGSYMBOLICWORLD MODELS FORSTOCHASTICENVIRONMENTS FROM UNGUIDEDEXPLORATION.
