GoldenRetriever Export Catalog

This page mirrors the live [tool.retriever.module.exports] manifest in pyproject.toml. Loading any of these must not require a camera, simulator, model, robot, network service, or source-tree knowledge — they resolve to plain payload classes and two conversion helpers.

The manifest

Export	Source	First use
`WorldState`	`retriever_typing.robotics_types:WorldState`	Perception, memory, planning.
`RobotState`	`retriever_typing.robotics_types:RobotState`	Robot wrapper, monitor, policy.
`BeliefGraph`	`retriever_typing.robotics_types:BeliefGraph`	Memory and grounding.
`Skill`	`retriever_typing.robotics_types:Skill`	Planner-to-controller handoff.
`Plan`	`retriever_typing.robotics_types:Plan`	Task planning.
`StructuredPlan`	`retriever_typing.robotics_types:StructuredPlan`	Typed multi-step planning.
`TaskGoal`	`retriever_typing.robotics_types:TaskGoal`	Intent, affordances, success criteria.
`Trajectory`	`retriever_typing.robotics_types:Trajectory`	Motion planner / controller.
`ExecutionStatus`	`retriever_typing.robotics_types:ExecutionStatus`	Monitor and controller outputs.
`Action`	`retriever_typing.core_types:Action`	Lightweight action payloads.
`Command`	`retriever_typing.core_types:Command`	Runtime or robot command payloads.
`Status`	`retriever_typing.core_types:Status`	Status exchange and smoke demos.
`convert_to_arrow`	`retriever_typing.conversions:convert_to_arrow`	Dataset / event-stream serialization.
`convert_from_arrow`	`retriever_typing.conversions:convert_from_arrow`	Dataset / event-stream deserialization.

By job

Scene / memory — WorldState, RobotState, BeliefGraph.
Task / plan — TaskGoal, Skill, Plan, StructuredPlan.
Control — Trajectory, ExecutionStatus.
Interop — Action, Command, Status, convert_to_arrow, convert_from_arrow.

Use it directly, or through Hub

In the GoldenRetriever source tree, import directly. From another project (once the repo and Hub index are public), the same names load through Hub and resolve to the same classes:

from retriever import hub
from retriever_typing.robotics_types import WorldState

# Once public — resolves to the manifest export above:
WorldStateFromHub = hub.use("openretriever/golden-retriever:WorldState")
assert WorldStateFromHub.__name__ == WorldState.__name__

The convert_to_arrow / convert_from_arrow pair is what makes the payloads dataset- and replay-friendly. The Hub smoke demo exercises exactly that on an exported Action:

pixi run demo-golden-hub-pack

GoldenRetriever pack exports: WorldState, BeliefGraph, Skill, Plan, Trajectory, convert_to_arrow, convert_from_arrow
Registry WorldState: _retriever_hub.golden_hub_pack_smoke__retriever_typing.robotics_types.WorldState
Constructed WorldState: ['cup']
Constructed Plan skills: ['pick']
Arrow round-trip: Action OK
Hub reference: hub.use("openretriever/golden-retriever:WorldState")

Promotion rule: an export is public only when this catalog, the manifest, and demo-golden-hub-pack agree. If a payload needs an optional simulator/model/camera dependency to import, it stays source-only until it has a lightweight boundary.

Source: pyproject.toml [tool.retriever.module.exports]; demo examples/advanced/core_composition/golden_hub_pack_smoke.py.