ExoPredicator: Learning Abstract Models of Dynamic Worlds for Robot Planning

2025-09-30

AI Summary

Long-horizon robotic planning must account for exogenous environmental dynamics—such as water heating or domino effects—that evolve independently of the agent’s actions, yet existing world models often neglect such non-autonomous changes. This work introduces the first learnable abstract world model that jointly represents symbolic state abstractions and causally models both endogenous agent behaviors and exogenous environmental mechanisms. Methodologically, it integrates large language model (LLM)-guided structural priors with variational Bayesian inference to efficiently learn stochastic temporal causal effects from few-shot observations. Its key contribution is the first explicit incorporation of exogenous processes into a learnable symbolic causal model, enabling unified representation and reasoning over both endogenous and exogenous dynamics. Evaluated on five simulated robotic tasks, the model achieves significantly higher long-horizon planning success rates and superior cross-task generalization compared to multiple baselines.

Technology Category

• Humans and AI: Human-Aware Planning and Behavior Prediction
• Planning, Routing, and Scheduling: Model-Based Reasoning
• Machine Learning: Causal Learning

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Long-horizon embodied planning is challenging because the world does not only change through an agent's actions: exogenous processes (e.g., water heating, dominoes cascading) unfold concurrently with the agent's actions. We propose a framework for abstract world models that jointly learns (i) symbolic state representations and (ii) causal processes for both endogenous actions and exogenous mechanisms. Each causal process models the time course of a stochastic causal-effect relation. We learn these world models from limited data via variational Bayesian inference combined with LLM proposals. Across five simulated tabletop robotics environments, the learned models enable fast planning that generalizes to held-out tasks with more objects and more complex goals, outperforming a range of baselines.

Problem

• Modeling dynamic worlds with concurrent endogenous and exogenous processes
• Learning symbolic states and causal processes from limited data
• Enabling fast planning generalization in complex robotics environments

Innovation

• Learns symbolic states and causal processes
• Uses variational Bayesian inference with LLM proposals
• Enables fast planning in complex simulated environments