Abstract
The ability of artificial intelligence agents to make optimal decisions andgeneralise them to different domains and tasks is compromised in complexscenarios. One way to address this issue has focused on learning efficientrepresentations of the world and on how the actions of agents affect them instate-action transitions. Whereas such representations are procedurallyefficient, they lack structural richness. To address this problem, we proposeto enhance the agent's ontology and extend the traditional conceptualisation oftrajectories to provide a more nuanced view of task execution. StructurallyEnriched Trajectories (SETs) extend the encoding of sequences of states andtheir transitions by incorporating hierarchical relations between objects,interactions, and affordances. SETs are built as multi-level graphs, providinga detailed representation of the agent dynamics and a transferable functionalabstraction of the task. SETs are integrated into an architecture, StructurallyEnriched Trajectory Learning and Encoding (SETLE), that employs a heterogeneousgraph-based memory structure of multi-level relational dependencies essentialfor generalisation. We demonstrate that SETLE can support downstream tasks,enabling agents to recognise task relevant structural patterns across CREATEand MiniGrid environments. Finally, we integrate SETLE with reinforcementlearning and show measurable improvements in downstream performance, includingbreakthrough success rates in complex, sparse-reward tasks.