Skip to main content

Introduce Persistent Rules

A persistent rule has a construction-side effect and an interaction-side presentation. Keep those two parts consistent.

A symbolic assignment​

from magma_core.simulation.state import TaskState
from magma_scenarios.templates.constraints import RelationAssignmentConstraint

state = TaskState()
state.attributes = {"objects": ["cup"], "target_areas": ["zone_a", "zone_b"]}
constraint = RelationAssignmentConstraint(
    "cup", "zone_a", "object_area",
    source_attribute_key="objects",
    target_attribute_key="target_areas",
)
constraint.apply(state)

This records the assignment and constraint history. It does not move the cup or send a message to the agent.

Present the rule​

from magma_core.simulation.stage import ConstraintBaseStage

stage = ConstraintBaseStage("From now on, place the cup in zone_a.", reset_at_end=False)

This checks the response to the new rule. A later action stage should use the assignment to construct the expected goal. Acknowledging a rule and complying with it are separate checkpoints.

In a request, sample the assignment once, construct the rule stage from it, and apply the corresponding constraint in apply_request. BaseConstraintRequest with ConstraintParameters implements that common split, but concrete requests still provide parameter sampling.

Attribute changes and replay​

When a referenced object or area disappears from the symbolic vocabulary, constraints can become outdated. Requests that modify attributes can request state replay. Preserve constraints in history and implement outdated correctly rather than relying on unrecorded relation mutations.

Next: Modify attributes. See requests and constraints for the full construction cycle.