Robots
Foundation-model robotics will reach AU logistics, retail and mining as a data and integration problem before it reaches them as a hardware problem; the firm's relevance is the telemetry and decision layer, not the robot.
Clustered only. No lab work behind it. Cannot be cited.
Confidence
38%unresearchedExpiry
34doverdue for reviewLead time
−6moopened after mainstream — recorded honestlyOwnership
Unownedcandidate — a named human electsWhere it is
Vision-language-action models made general-purpose manipulation demos routine in 2026 and made none of them dependable. Published task success rates on novel objects sit around 70–85% in the lab; a warehouse needs four nines on the tasks it already automates. The economic reality in AU is that Woolworths' automated DCs, the Pilbara's autonomous haul fleets and Amazon's fulfilment robots all run on narrow, non-foundation systems that work. Where foundation models enter is the long tail: exception handling, novel-object picking, natural-language tasking. That is a software layer sitting on telemetry and decision data, which is closer to what Quantium does than the hardware is.
Why a Quantium decision hinges on it
Woolworths is the owner and the largest AU operator of automated retail logistics. If foundation-model robotics changes the economics of the long tail — the 15% of picks that still go to humans — the decision data around it is a Quantium-shaped problem. The firm does not build robots and should not; it should know when the layer above the robot becomes a modelling problem.
Field attributes
Position
What is demonstrated, what is hype, what would have to be true.
The shape every position request answers. Signal-tier fields carry a draft; assessed and tested fields carry a validated one.
- 01Open-weight VLA models perform novel-object pick-and-place at 70–85% success in controlled settings across several robot bodies.
- 02Narrow automation in AU logistics and mining is mature, profitable and not foundation-model based.
- 03Natural-language tasking of a warehouse robot works in demos; no published deployment reports exception rates.
- 01Humanoid form factor. The body is the story and the story is not the economics.
- 02Demo success rates on curated object sets presented as deployment readiness.
- 03'General-purpose' as a claim about a model that has been trained on one lab's robots.
- 01A published deployment where a foundation-model policy handles the long-tail exceptions in a live DC at a success rate the operator accepts, with the exception rate reported.
- 02Sim-to-real transfer from a learned world model to a real robot on a task we could specify, at a published success rate.
- 03A telemetry and decision layer that the operator will let a third party model — the Quantium entry point.
- 01Nothing hardware-side, ever. The field is watched for the software layer.
- 02Ask Woolworths' DC automation team what their long-tail exception rate is and whether they log it; that number is the demand signal we lack.
- 03If the deployment trigger fires: Type 3 survey run into the decision-data layer around a live deployment, with the retail sector owner.
Signals · 10 in this cluster
What the cluster is made of.
Every item carries its source, tier and sightings. Detector-found signal sits beside human drops; downstream they are indistinguishable except by provenance.
Open cross-embodiment manipulation benchmark: best VLA at 81% on novel objects
Standardised novel-object pick-and-place across six robot bodies. Best open-weight policy 81%, best closed 84%; humans 99%+. The gap to deployment is this number, and it is public.
extracted claimFoundation-model manipulation sits at 70–85% success on novel objects.
'Can you model which exceptions we should automate next?'
Asked by a supply-chain analytics lead on the owner's side. Not a robotics question; a prioritisation question with robot options. Logged as the field's one real demand signal.
Training Manipulation Policies Inside Learned World Models: Sim-to-Real at Scale
Policies trained mostly in a learned simulator, fine-tuned on 200 real demos, reach 76% on the cross-embodiment benchmark. Not yet parity with real-data training but a 10× cut in real-data collection. The bridge from World Models to this field.
'The last 15%: what still goes to humans in an automated DC'
An automation lead at a large AU grocer's DC describes the exception tail: damaged packaging, novel SKUs, mixed pallets. Says the number is logged but never modelled. This is the closest thing to a demand signal the field has.
extracted claimThe long tail of DC exceptions is logged, unmodelled, and is where foundation models would enter.
'The humanoid is the wrong shape for every job except the demo'
Practitioner argument that wheeled, fixed and gantry systems beat humanoids on cost per pick and always will inside a building designed for them. Widely shared among operators. The counter-signal to the humanoid press cycle.
Open-weight VLA model released with fine-tuning recipe for new robot bodies
First open-weight policy with a documented recipe for adapting to a new body in under a thousand demonstrations. Makes the field reproducible outside the labs that own robots.
extracted claimAdapting a foundation policy to a new robot body now needs hundreds, not tens of thousands, of demonstrations.
Pilbara operator hiring 'Autonomy Decision Analytics Lead' ×2
Two roles for analytics over autonomous fleet data. Argus inference: the operator agrees the next decade is decisions, not driving, and is staffing it in-house rather than through a consultancy.
Humanoid vendor announces retail-store pilot 'shelf restocking' with a US grocer
Pilot announcement with no success rate, no throughput and no cost. Widely covered. Kept as the humanoid claim; nothing in the release supports store economics.
'Autonomous haulage is done; the next decade is maintenance'
Pilbara operator's autonomy lead: driving is solved with narrow systems; the open problems are maintenance scheduling, inspection and exception decisions. Foundation models enter as analytics, not as drivers.
openfleet — fleet exception analytics for mixed robot estates
Open-source exception-logging and analytics layer for heterogeneous robot fleets. Low stars, but the existence of an unbundled layer is evidence against the red team's bundling case.
Claims · 4 supporting, 1 refuting
The atoms.
A document cannot go stale; an assertion can. Claims are immutable and stamped with the extractor that produced them, so staleness, diffs and the graveyard operate at claim level.
Foundation-model manipulation policies sit at 70–85% task success on novel objects; deployment needs above 99% on the tasks currently automated.
Mining autonomy in the Pilbara is mature and narrow; foundation models enter through maintenance and exception decisions, not driving.
The economic entry point in AU logistics is the long tail of exceptions, which is a decision-data problem before it is a hardware one.
Sim-to-real from learned world models will shorten the data-collection bottleneck for new tasks by an order of magnitude.
Humanoid robots will be economically deployed in retail stores within the Next horizon.
Position history · the diff is the product
2 validation runs against a fixed brief. Confidence 35% → 38%.
Long-tail exceptions named as the entry point. Humanoid claims down-weighted. Red team's bundling argument is unanswered and the review date has now lapsed.
- The economic entry point in AU logistics is the long tail of exceptions, which is a decision-data problem before it is a hardware one.
- Mining autonomy in the Pilbara is mature and narrow; foundation models enter through maintenance and exception decisions, not driving.
- Sim-to-real from learned world models will shorten the data-collection bottleneck for new tasks by an order of magnitude.
- c-robots-3 ↓ 0.26 → 0.18
Scoring · ordinal bands
Agents propose. A named human commits.
Uncommitted scores are visibly marked and never leave the building. Bands, not point estimates — false precision is the tell that a number was generated rather than derived.
Impact
committed · DSReal for the owner's logistics arm; the firm's slice is the decision layer, which is medium at best.
Timeline
agent-estimatedAgent-estimated: long-tail deployment in a DC within four years is plausible on the reliability trajectory. Uncommitted.
TAM
agent-estimatedAgent-estimated on global logistics robotics; the addressable decision-layer slice is unknown.
Cost
committed · TOWatching costs nothing; a survey run into a live deployment is a Type 3.
Cost of being wrong
committed · LFThe firm is not building hardware; a late start on the software layer costs an engagement, not a business.
Demand
committed · DSOne question from the owner's supply-chain side about exception handling; nothing from other retail clients.
Cost of entry
agent-estimatedDomain credibility with DC operators has to be earned; the firm has none in robotics. Agent-estimated.
Relevance · per vertical
Why it matters here, or explicitly does not.
Ranking is per vertical, not global. Sector owners commit notes against agent drafts.
The owner runs AU's largest automated retail logistics estate. The long-tail exception layer is where a foundation-model policy would enter and where the decision data lives.
Mechanism · Exception-rate telemetry from automated DCs modelled as a decision problem; a foundation-model policy as one option in the decision.
Mining and utilities run mature narrow autonomy. Foundation models enter through maintenance and inspection decisions, which are analytics problems.
Mechanism · Inspection-robot telemetry as input to asset-health models.
Surgical and care robotics are a regulatory and hardware domain the firm has no path into on this horizon.
Mechanism · None.
Autonomy is the fastest-moving defence procurement category and the firm has no presence; a sponsor question, not an owner one.
Mechanism · Would require a sponsor to hold the 'should we be here' question.
Red team · the strongest case against
The strongest case against: our 'software layer' framing is a consultancy telling itself the hardware transition is someone else's problem. The operators who deploy foundation-model robots will get the decision layer from the robot vendor, not from an analytics firm. Our entry point may not exist.
- —Robot vendors ship the fleet-management and exception-analytics layer bundled; there is no unbundled decision problem for a third party to model.
- —70–85% on novel objects was 40% two years ago; the reliability trajectory is steeper than our timeline band assumes.
- —The owner's DC automation is run by an integrator under a long contract; the firm's ownership link is not an access link.
Source diversity
- Robotics research35%
- AU operators / practitioners30%
- Vendor / analyst15%
- Internal / Engel20%
A field supported by one epistemic community is a flag, not a finding.
Cross-pollination · typed joins
Connected, not merely similar.
Enabling, compounding, substituting, blocking. A satisfied dependency trigger is a far stronger signal than semantic proximity.
Trigger · A published sim-to-real result where a policy trained in a learned world model reaches ≥90% success on a real manipulation task the lab can specify, on a robot body available in AU; Argus tracks the three VLA labs' release notes for it.
When the trigger fires, this field is resurfaced automatically. Watchable rather than parked.
Trigger · A DC operator (the owner's or a competitor's) publishes or shares a long-tail exception rate from a live foundation-model deployment; until an exception rate exists there is no decision problem to model.
When the trigger fires, this field is resurfaced automatically. Watchable rather than parked.
Learned simulators are the training environment and the sim-to-real path for new tasks.
Robot fleet telemetry is the richest sensing source in a DC; the sensing agent pattern is the entry point.
Natural-language tasking of robots is a voice-and-vision interface problem.
The firm has no robotics credibility; the right to play in this room has not been bought.
Share graph
Provenance running forward.
Discovery, not accountability. No counts, no rankings, no rollups to managers.
Convergence · who else is here
- DSDev Sharma · Sector owner · Retail3 drops
- TOTom Okafor · Research engineer · agents2 drops
- MTMei Tanaka · Research lead · evals2 drops
- RMRohan Mehta · Exec sponsor1 drop
- ?Anonymous · Anonymous drop1 drop
Several people’s drops meet here. An informal working group already exists and probably does not know it.
Lineage
What this field produced, and what it killed.
Experiments, recommendations and graveyard entries stay attached. The reasoning that killed a claim is the reusable asset.
No experiments, recommendations or graveyard entries yet. That is what a candidate looks like.
Open questions · return to the pile
Every run leaves a record. Separately, its question either closes or returns to the pile with notes — which is what the next person proposing the same thing will see.
- 01Does an unbundled decision layer exist in a live DC, or do robot vendors ship it bundled?
- 02What is the owner's actual long-tail exception rate, and who owns the data?
- 03At what novel-object success rate does a DC operator let a foundation-model policy touch a live pick line?