Skip to content
GoldenRetrieverGoldenRetriever Reference

GoldenRetriever Hub Proof

The smallest check that a project outside the core runtime can publish reusable robot payloads through Retriever Hub. No robot, camera, simulator, model, network, or GUI is required.

pixi run demo-golden-hub-pack

What it does

Section titled “What it does”

The smoke test loads this repo’s [tool.retriever.module] manifest through the same runtime Hub loader that hub.use(...) uses, then exercises the exports: construct payloads, look one up in the unified registry, and round-trip a payload through the exported Arrow helpers.

import tomllib
from pathlib import Path
from retriever.hub._loader import load_exports
from retriever.registry.types import get_type_info, list_types

# 1. read the manifest and load exports through the Hub loader
manifest = tomllib.loads(Path("pyproject.toml").read_text())["tool"]["retriever"]["module"]
exports = load_exports(Path("."), manifest["module"], manifest["exports"],
                       namespace="golden_hub_pack_smoke",
                       hub_meta={"org": "openretriever", "name": "golden-retriever"})

WorldState, Skill, Plan, Action = (exports[n] for n in ("WorldState", "Skill", "Plan", "Action"))
convert_to_arrow, convert_from_arrow = exports["convert_to_arrow"], exports["convert_from_arrow"]

# 2. construct robot-facing payloads
world = WorldState(object_poses={"cup": pose}, robot_pose=pose, timestamp=0.0)
plan  = Plan(skills=[Skill(name="pick", params={"object": "cup"}, confidence=0.95)])
action = Action(type="pick", parameters={"object": "cup"}, timestamp=0.0, priority=1)

# 3. Arrow round-trip must be lossless
assert convert_from_arrow(convert_to_arrow(action), Action) == action

# 4. the type is visible in the unified registry
assert "WorldState" in list_types()
print("Registry WorldState:", get_type_info("WorldState").type_class.__module__)

What you get

Section titled “What you get”

Real output. The exports load, WorldState resolves through the registry under a Hub-namespaced module, payloads construct, and the Arrow round-trip is lossless:

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")
Graph proof: run `pixi run demo-pipeline-html-viz` to validate and render an IR HTML artifact.

Hub reference shape

Section titled “Hub reference shape”

Once the repo and Hub index are public, the same payloads load over Hub with no manifest plumbing — install the runtime once, then pull domain packs by name:

from retriever import hub

WorldState       = hub.use("openretriever/golden-retriever:WorldState")
Plan             = hub.use("openretriever/golden-retriever:Plan")
convert_to_arrow = hub.use("openretriever/golden-retriever:convert_to_arrow")

Why it’s shaped this way

Section titled “Why it’s shaped this way”

Loading through the manifest (not a direct import) is the point: it proves the pack boundary the same way a remote Hub fetch would, while staying offline. The registry lookup and Arrow round-trip confirm the exports are real runtime types — usable and serializable — not just names in a file.

Continue with the export catalog for the exact exports and their maturity, or the example catalog for the applied lanes.

Source: examples/advanced/core_composition/golden_hub_pack_smoke.py.