13.4 Schema learning in an embodied text world

The schema experiments move from abstract switching processes to ALFWorld, an environment that aligns textual interaction with embodied household tasks [ Shridhar et al. , 2021 ] . An agent may need to find an object, acquire it, operate on it, transport it, and place it in a receptacle. Observations are partial, object locations persist across time, and action validity depends on relations such as containment, co-location, carrying capacity, receptacle state, and tool compatibility.

This setting reconnects RELIC with the Piaget–Drescher lineage. A schema is not merely a label attached to a trajectory. It predicts which action is applicable, what transition should follow, and which persistent relations must be updated [ Drescher , 1991 ] . Failure creates a counter-witness to the current schema; repeated evidence can justify accommodation.

The observation-only studies recovered increasingly relational structure:

• action-family success conditions were learned from textual outcomes;

• source, destination, tool, and receptacle types were separated;

• object containment was represented as a changing location relation;

• acquisition required both object–source and agent–source co-location;

• carrying capacity and persistent tool location were added; and

• frozen schemas were replayed on disjoint games before being used by a planner.

This sequence matters because a strong frozen policy could otherwise conceal the origin of its structure. Every retained predicate has an evidence record and a counter-witness history. Still, the final artifact and the process constructing it must be evaluated separately.