OpenNovelty.org

ExoPredicator: Learning Abstract Models of Dynamic Worlds for Robot Planning

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.5Download Report PDF
learning abstractions for planningneuro-symbolic aiconcept learning
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.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper proposes a framework for learning abstract world models that jointly represent symbolic states and causal processes for both endogenous actions and exogenous mechanisms, enabling long-horizon robot planning in environments where external processes unfold concurrently with agent actions. Within the taxonomy, it occupies the 'Abstract World Models with Exogenous Dynamics' leaf under 'World Model Learning and State Abstraction'. Notably, this leaf contains only the original paper itself—no sibling papers are present—indicating this is a relatively sparse and potentially underexplored research direction within the broader field of 50 papers surveyed.

The taxonomy reveals that the paper's parent branch, 'World Model Learning and State Abstraction', contains two neighboring leaves: 'Latent State Discovery and Control-Endogenous Representations' and 'State and Action Abstraction for Planning'. These adjacent directions focus on filtering task-relevant information and hierarchical abstractions respectively, but explicitly exclude causal modeling of exogenous processes (per the exclude_note). The broader field shows substantial activity in MPC-based methods (25 papers across six leaves) and learning-based control (5 papers), suggesting the paper diverges from dominant optimization-centric and purely data-driven paradigms by emphasizing symbolic causal reasoning over external dynamics.

Among 28 candidates examined across three contributions, none were found to clearly refute any claimed novelty. The 'Framework for abstract world models with exogenous processes' examined 10 candidates with 0 refutable; 'Variational Bayesian inference method for learning causal models' examined 10 with 0 refutable; and 'State abstraction learner using foundation models' examined 8 with 0 refutable. This suggests that within the limited search scope, the combination of symbolic causal modeling for exogenous dynamics, variational inference for learning, and LLM-based state abstraction appears relatively unexplored in the examined literature, though the search scale (28 papers) is modest relative to the field's breadth.

Based on the top-28 semantic matches and taxonomy structure, the work appears to occupy a distinct niche: explicitly modeling exogenous causal processes in abstract world models for robotics. The absence of sibling papers in its taxonomy leaf and the lack of refuting prior work among examined candidates suggest potential novelty, though the limited search scope means more comprehensive surveys or domain-specific venues might reveal closer precedents. The analysis covers semantic similarity and citation-based expansion but does not exhaustively survey all world modeling or causal inference literature in robotics.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
28
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: learning abstract world models for long-horizon robot planning with exogenous dynamics. The field addresses how robots can plan and act over extended horizons when faced with external disturbances or unmodeled forces. The taxonomy reveals a rich landscape organized around nine major branches. Model Predictive Control for Robot Manipulation and Locomotion (e.g., Data-driven MPC Trajectory[1], Contact-Implicit Dexterous MPC[4]) focuses on optimization-based methods that handle contact dynamics and terrain variations. Vision-Language and LLM-Based Planning (e.g., Vision-Language MPC[2], LLM Deformable Assembly[3]) leverages modern foundation models to ground high-level instructions in robotic actions. Learning-Based Control and Reinforcement Learning emphasizes data-driven policy synthesis, while World Model Learning and State Abstraction tackles the challenge of building compact, predictive representations of environment dynamics. Sim-to-Real Transfer and Multi-Environment Learning addresses generalization across diverse settings, and Reactive Planning and Perturbation Recovery (e.g., Reactive Locomotion Perturbation[31]) deals with online adaptation to unexpected disturbances. The remaining branches cover Human-Robot Interaction, Multi-Agent Coordination, and Specialized Applications, reflecting the breadth of robotic planning challenges. A central tension across these branches is the trade-off between model fidelity and computational tractability: detailed physics-based models (as in Contact-Implicit Dexterous MPC[4] or Quadruped Slippery Surfaces[5]) offer precision but can be expensive, whereas learned abstractions promise efficiency at the cost of potential modeling errors. Within World Model Learning and State Abstraction, works like Control-Endogenous Latent States[21] and State Action Abstractions[20] explore how to distill high-dimensional observations into compact latent dynamics. ExoPredicator[0] sits squarely in this branch, emphasizing abstract representations that explicitly account for exogenous factors—external influences not directly controlled by the robot—enabling more robust long-horizon planning. Compared to approaches that assume fully observable or control-endogenous dynamics (e.g., Control-Endogenous Latent States[21]), ExoPredicator[0] addresses scenarios where environmental changes arise independently, a distinction that becomes critical in real-world settings with unpredictable disturbances or multi-agent interactions.

Claimed Contributions

Framework for abstract world models with exogenous processes

The authors introduce a framework that learns symbolic state abstractions and causal processes modeling both agent actions (endogenous) and external environmental dynamics (exogenous) that unfold concurrently with agent actions, enabling abstraction over temporal granularity.

10 retrieved papers
Variational Bayesian inference method for learning causal models

The paper contributes an efficient Bayesian inference method that learns the parameters and structures of causal processes from limited trajectory data, using variational inference for continuous parameters and LLM-guided proposals for discrete structure search.

10 retrieved papers
State abstraction learner using foundation models

The authors develop a method for learning symbolic state abstractions (predicates) by prompting language models to propose candidate predicates and then performing local search to select subsets that optimize Bayesian objectives.

8 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Within the taxonomy built over the current TopK core-task papers, the original paper is assigned to a leaf with no direct siblings and no cousin branches under the same grandparent topic. In this retrieved landscape, it appears structurally isolated, which is one partial signal of novelty, but still constrained by search coverage and taxonomy granularity.

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Framework for abstract world models with exogenous processes

The authors introduce a framework that learns symbolic state abstractions and causal processes modeling both agent actions (endogenous) and external environmental dynamics (exogenous) that unfold concurrently with agent actions, enabling abstraction over temporal granularity.

Contribution

Variational Bayesian inference method for learning causal models

The paper contributes an efficient Bayesian inference method that learns the parameters and structures of causal processes from limited trajectory data, using variational inference for continuous parameters and LLM-guided proposals for discrete structure search.

Contribution

State abstraction learner using foundation models

The authors develop a method for learning symbolic state abstractions (predicates) by prompting language models to propose candidate predicates and then performing local search to select subsets that optimize Bayesian objectives.

Table of Contents