Included
- Customer-facing robot picking AI demo
- Grasp candidate scoring
- Failure-risk explanation
- Local-first pilot blueprint
Robotic World Model
ARK NOVA Pick
Local-first robot picking AI that predicts failure before robot motion.
A local-first world model MVP for warehouse automation and robot picking AI. It selects the target object, ranks grasp candidates, predicts slip, collision, occlusion, and unknown-SKU risk, then recommends the safest local action before any robot moves.
A low-cost path from public synthetic robot picking AI demo to paid warehouse automation pilot without building custom robot hardware first.
Local runtimeWorld modelRobot picking AIWarehouse automationRisk scoring
Pilot thesis
Sellable pilot unit
A warehouse buyer can start with one local decision loop: camera scene in, ranked pick plan out, failure risks explained before a robot moves.
First buyerWarehouse picking teams
Pilot scopeOne bin, one camera, one robot-cell workflow
Runtime stanceNo cloud dependency
- Reduce failed pick trials by scoring slip, collision, occlusion, and unknown-SKU risk before motion.
- Keep customer footage, SKU labels, safety rules, and robot logs inside the buyer environment.
- Move from public synthetic demo to a scoped paid pilot without waiting for custom robot hardware.
Customization path
- Tune to customer SKU sets
- Connect to customer robot arms through ROS2
- Move from synthetic demo to edge runtime for pilots
Best buyer
Warehouses, small fulfillment teams, and automation partners testing robotic picking.
Try the local flow
The public demo shows a synthetic warehouse bin scene, runs the local-first world model pipeline, and explains why the robot picking AI chooses pick, skip, ask human, or reposition camera.
$ Open the Hugging Face public demo$ Choose a warehouse picking scenario$ Run the decision loop$ Review target, grasp candidates, and risk cardsExpected demo output
The demo proves the first sellable unit: local target selection, grasp ranking, risk prediction, and action recommendation before movement.
Recommended action: PICK / ASK HUMAN / REPOSITION CAMERAPredicted success and failure risks before motionRanked grasp candidates with explanationDeveloper JSON view for auditabilityPurchase delivery
- Public demo for investor and customer validation
- GitHub pilot repo for technical diligence
- Customer pilot scope: SKU set, camera setup, robot interface, local runtime
What changes for the buyer
For a warehouse automation pilot, the synthetic scene is replaced with the buyer's own SKU set, camera geometry, robot arm constraints, local logs, and approval workflow. Customer data stays private.
Buyer clarity
Before purchase
Clear answers for delivery, customization, and license boundaries.
Is this a robot hardware product?
No. ARK NOVA Pick is a robot intelligence layer. The public Space is only a showroom; the real product direction is local-first and edge-deployable.
Does the production product require cloud inference?
No. The operating principle is no cloud dependency. Cloud is optional only for public demos, approved reporting, support bundles, or distribution.
What is proven today?
MVP-1 proves the decision loop: target selection, grasp candidate generation, failure-risk scoring, and recommended action before robot movement.
Product proof
Public Gradio demo for robot picking AI is live on Hugging Face.
Product proof
GitHub repo contains the local decision loop, tests, docs, and pilot plan.
Product proof
The product is positioned as no cloud dependency: customer footage, SKU data, robot control, and safety rules stay local/private.
