Modeliong Uncertainty in World Models

Riding the wave of recent advances in vision and language foundation models and video generation models~\cite{wan_Wan_2025}, learning world models to fully capture diverse environments has become an active area of research~\cite{wu_oscar_2026,tong_pointaction_2026,li_Turning_2026,guo_Unified_2026}. World models can be used to generate task plans, simulate policies, and be used as reward models to score trajectories for robotics. World models enable robot decision-making to be less reliant on hand-crafted simulators and real-world rollouts.

However, robots operating in the real world face uncertainty that most work in world models fail to address. The sensors can be noisy, environments can be partially observable, there can be observation and transition noise, and imprecise actuators can all affect robot behavior~\cite{wang_ManiDreams_2026,singh_Structured_2021,saleem_POMDPbased_2024}. Many environments also feature humans, robots, or other agents that introduce additional uncertainty about their goals and intent that is difficult but important for robots to anticipate~\cite{liang_Introspective_2024,leusmann_Uncertainty_2023,ma_Flexible_2026}. Incorporating uncertainty into robotic world models raises a set of open technical challenges. Prior work has applied belief-space planning methods~\cite{garrett_Integrated_2020,fang_Embodied_2024a,saleem_POMDPbased_2024}, conformal prediction~\cite{liang_Introspective_2024}, and goal inference~\cite{ma_Flexible_2026} to address uncertainty in other apsects of robotics.

While generative world models in their current form may be new, incorporating uncertainty that commonly arises in partially observable systems has been widely studied in various contexts, e.g., POMDPs in robotics, automated planning, and computational learning theory; intuitive physics in cognitive science; and learning sequential or recurrent models in machine learning. Through invited talks, interactive poster sessions, and panel discussions, we hope to arrive at an understanding of how the uses of world models in different robotic domains inform the types of uncertainty we should incorporate.

This workshop brings together researchers from robotic planning, probabilistic robotics, safe autonomy, uncertainty quantification, and human-robot interaction to directly confront these challenges. We aim to clarify which uncertainty representations (distributional, latent, conformal, or ensemble-based) are useful suited to which robotic settings and tasks, how uncertainty propagates from world models into downstream planning and reinforcement learning, and what evaluation benchmarks and metrics are needed to measure progress. Through invited talks, interactive poster sessions, and interactive roundtable discussions, our goal is to encourage the development of world models that are not merely expressive, but reliably uncertainty-aware, which is a necessary step toward robust robot deployment in unstructured real-world environments.

Core Challenges / Research Questions:

We aim to address several core research challenges and propose open questions in two main groups centered around uncertainty-aware robotic world models:

Importance of Modeling Uncertainty in Robotic World Models

• What are definitions of robotic world models, and why should we model uncertainties in robotic world models?

• Are there certain types of uncertainties more important to model in robotic world models?

• How do we consider uncertainty with modelling other robots (in multi-agent settings) or humans in the environment?

• Are there computational tractability issues that arise when we allow for certain types of uncertainty when learning a world model?

• How does uncertainty modeling affect the generation quality of robotic world models?

How to Model Uncertainty for Robotic World Models

• How do different definitions and implementations of robotic world models affect how uncertaity should be modelled?

• How do we model different types of uncertainty, i.e., epistemic and aleatoric, in robotic world models? Should we model aleatoric uncertainty and epistemic uncertainty jointly or separately?

•  How do we account for uncertainties in the robotic world model in a computationally tractable way? 

•  What is the applicability of robotic world models in perception under uncertainty, belief-space planning, model-based RL, human-aware interactions, and multi-agent coordination?

•  Can we and how do we collect sufficient data to train robotic world models to be robust to uncertainty?

•  We have many techniques to model uncertainty, e.g., conformal methods, latent representation learning methods, noise models, out-of-distribution detection, etc. How do we justify that a technique is more appropriate over the others for what problem? Empirically, how well do different techniques work for different problems?

• How do uncertainties incorporated into world models influence downstream processes that leverage the world model to determine robot behavior, e.g., reinforcement learning, planning, etc.?

• What domains and evaluation metrics can we use to evaluate robotic world models?