Semantic World Models for Reasoning and Decision-Making in Uncertain Environments

Semantic mapping refers to the process of constructing a representation of the environment that not only captures spatial information but also encodes the meaning of objects and regions within it. In this framework, elements of the environment are assigned semantic labels—for example, a robot may classify a region as traversable terrain, obstacle, or restricted area. This enables the robot to reason about its surroundings rather than simply reacting to raw sensor data. However, current approaches are often limited by their reliance on predefined datasets and static models. As a result, adapting to new environments or unseen conditions frequently requires retraining or redesigning the system. This limitation becomes particularly significant in complex scenarios where multiple factors—such as terrain type, environmental conditions, and task constraints—interact in non-trivial ways.

At CAST, we advance semantic mapping by using conventional artificial intelligence approaches with domain knowledge to model complex environments more effectively. In particular, we leverage description logics and ontology-based representations to encode relationships between entities in the environment, enabling robots to move beyond simple perception toward meaningful reasoning.By incorporating these semantic models, robots are able to make knowledge-driven inferences based on defined rules, rather than relying solely on pre-trained datasets. This allows systems to adapt more effectively to new and previously unseen environments without requiring complete retraining. Furthermore, we are working towards how semantic mapping can be tightly coupled with uncertainty-aware decision-making frameworks.