Sector · Warehousing and intralogistics
Physical AI for warehousing and intralogistics
These are the questions engineers ask first, answered from 1 published CodeNinja Atoms reference architecture.
- What does a physical AI system for warehousing and intralogistics look like?
- Which AI models can a warehousing and intralogistics operator run on its own hardware?
- How much compute and hardware does AI in warehousing and intralogistics need?
- Is it cheaper to own AI hardware or rent cloud GPUs in warehousing and intralogistics?
- What ontology or object model does a warehousing and intralogistics AI system need?
- Who approves the decisions an AI system makes in warehousing and intralogistics?
What does a physical AI system for warehousing and intralogistics look like?
A complete physical AI design for warehousing and intralogistics names what to sense, which existing systems to join, the object model that joins them, the models and hardware, the three-year cost and the person who approves every action. CodeNinja Atoms has published 1 such reference architecture for warehousing and intralogistics, each free to reuse under CC BY 4.0.
| Design | Country | What it does |
|---|---|---|
| Warehouse Twin Command | United States | An open reference architecture for system design in physical AI. A live digital twin of four distribution centers run by a warehousing and intralogistics operator in the United States: one object model projected over the warehouse, labor and yard systems the operation already runs, kept true by nightly reconciliation against the WMS, producing weekly slotting recommendations, dock and yard schedules and on-site safety vision on the existing cameras, with every action drafted for a person to approve. |
Which AI models can a warehousing and intralogistics operator run on its own hardware?
Each published warehousing and intralogistics design names its models and why, and every model is open-weight or no model is used at all, so the operator can run it on hardware it owns.
| Design | Models |
|---|---|
| Warehouse Twin Command | Mistral Large 3 675B (Apache-2.0) for the site manager assistant; Qwen3.8-27B (Apache-2.0) for assistant continuity at each site; RF-DETR N to L (Apache-2.0) and OC-SORT (MIT) for safety vision; BGE-M3 (MIT) for retrieval; no face recognition, no re-identification, no training run |
| Design | Choice | What was picked | Why |
|---|---|---|---|
| Warehouse Twin Command | Frontier model | Mistral Large 3 675B, Apache-2.0, FP8 | No revenue trigger or attribution duty; one 141 GB HBM class node holds the 682 GB weights with about 350 GB KV ceiling |
| Warehouse Twin Command | Site model | Qwen3.8-27B, Apache-2.0, one 48 GB PCIe card | About 31 GB at FP8 keeps the assistant alive in degraded mode during WAN loss |
| Warehouse Twin Command | Detection model | RF-DETR, Apache-2.0 N to L checkpoints | XL and 2XL carry a platform license; N to L takes site fine-tuning for evaluation sets |
| Warehouse Twin Command | Tracking model | OC-SORT, MIT, motion-only | No re-identification branch, so the safety system never becomes a surveillance system |
| Warehouse Twin Command | Embedding model | BGE-M3, MIT, 8,192 token window | One pass yields dense, sparse and multi-vector retrieval on a small box |
| Warehouse Twin Command | Edge compute class | Rugged rack server with PCIe inference accelerators | Sized per site from the measured decode-plus-inference budget on camera substreams, with headroom on both |
| Warehouse Twin Command | Camera estate | The existing roughly 220 fixed IP cameras, reused | Six-gate survey per camera; purpose-placed fixed additions only where a view fails |
| Warehouse Twin Command | Sizing rules | Streams, calibration, meters, bandwidth | Size from streams not boxes; calibrate when speeds must survive challenge; budget video on the correlated peak |
| Warehouse Twin Command | Sensing | WMS event stream, labor extracts, yard appointments, link and GPU health telemetry | The twin's feeds and the failure signals the boundary behavior depends on |
| Warehouse Twin Command | Patterns | System of Context, camera analytics as feedback loop, baseline before anything, shadow mode | The assistant drafts and a person approves; analytics never act as a safety-rated stop |
| Warehouse Twin Command | Ground | The operator's own hardware, open weights, in-company environment | Domestic supply engages no export pathway; only events and short clips leave each building |
How much compute and hardware does AI in warehousing and intralogistics need?
The compute follows from the models: the published warehousing and intralogistics designs size it as follows, from no new hardware to a full GPU node.
| Design | Part | The design |
|---|---|---|
| Warehouse Twin Command | Compute | At each site, one industrial edge server with one 48 GB PCIe inference card running vision in Triton and the site model under vLLM; centrally, one 141 GB HBM class node of eight cards, on the operator's own hardware |
Is it cheaper to own AI hardware or rent cloud GPUs in warehousing and intralogistics?
Each warehousing and intralogistics design prices three years of ownership in its Appendix A, with every price cited, against renting the same capacity from a cloud region at its deepest three-year commitment where hardware is bought.
| Design | Line | Three years |
|---|---|---|
| Warehouse Twin Command | Three-year cost | About 597,000 US dollars to own the twelve cards (510,000 to 689,000) before the four edge chassis, which are priced by quotation; renting the same cards in AWS US East runs about 1.86 million on demand and about 699,000 on the deepest three-year commitment (Appendix A) |
What ontology or object model does a warehousing and intralogistics AI system need?
The object model is the part that makes the system an ontology rather than a pipeline: typed objects for the things in the warehousing and intralogistics operation, with properties, status values and typed links. Every published design ships its object model as hyper-ontology/1 JSON that loads into Hyper Ontology.
| Design | Size | Objects | Download |
|---|---|---|---|
| Warehouse Twin Command | 14 objects, 14 links | Distribution Center, Dock Door, Yard Trailer, Storage Slot, SKU, Open Order, Pick Task, Associate, Labor Shift, Camera Zone, Safety Event, Forklift, Shift Lead, Slotting Recommendation | objects.json |
Who approves the decisions an AI system makes in warehousing and intralogistics?
In every published warehousing and intralogistics design a named person makes the decision that changes the physical world; the system prepares it.
| Design | Human control |
|---|---|
| Warehouse Twin Command | Every slotting move states walking saved and move labor before a person approves it; every safety event becomes a clip a supervisor reviews, and analytics never act as a safety-rated stop |
Every answer on this page is drawn from the papers linked in it: each paper's At a glance table, model register and cost appendix. Full text for agents: llms-full.txt. Designed on Praxis; object models load into Hyper Ontology.
Want one of these designs for your own operation? Talk to us.