HomePhysical AI › Reference Architectures
Physical AI · reference architectures

Robotics blueprints you can budget from

Validated physical-AI architectures — edge + manipulator cell, autonomous mobile robot, and VLA inspection cell. Each with a bill of materials, hardware options and indicative pricing. We tailor every one to your task.

These are starting points, not fixed products. Each blueprint pairs a model with the edge compute, sensors and actuator a task needs, with a bill of materials and indicative pricing you can plan around. We size every one to your environment and return a firm, tailored BOM — and because Haink builds the model and supplies the hardware, one partner owns the whole stack. Pricing is indicative, ex-works hub, and moves with robot and GPU allocation.

Edge Inference + Manipulator Cell

Pick-and-place / assembly1 robot armJetson AGX Orin / ThorVision-guided

The workhorse cell: a camera-guided robotic arm running a perception-plus-control or VLA policy on an edge node. Handles varied or unfamiliar parts where a fixed program would fail. Add IGX for functional safety where people are near the cell.

ComponentSpecificationNotes
Edge computeNVIDIA Jetson AGX Orin (or AGX Thor for VLA)Sized to model; Thor for large VLA
PerceptionIndustrial camera + depth sensorCalibrated to the workspace
Actuator6-axis cobot + gripperCollaborative arm with motion planning
ControllerMotion controller + safety I/OBounded autonomy, watchdogs
SoftwarePerception / control or VLA policyBuilt and optimized by Haink

~$40,000–$150,000 delivered cell incl. cobot arm, ex-works hub; arm alone is ~$25K–$60K in 2026

Autonomous Mobile Robot (AMR) Node

Warehouse / intralogisticsOn-board navigationLiDAR + camerasFleet-ready

The compute and perception stack for an autonomous mobile robot: on-board navigation and obstacle avoidance that keep running when connectivity drops, with telemetry and OTA updates for a fleet. Supplied as the AMR’s intelligence layer, integrated to your chassis or a new build.

ComponentSpecificationNotes
Edge computeNVIDIA Jetson AGX OrinNavigation + perception on-board
Perception2D/3D LiDAR + camerasMapping, localization, obstacle avoidance
ConnectivityWi-Fi 6 / 5GTelemetry and coordination
Fleet softwareMonitoring + OTA updatesOne robot or many
SoftwareNavigation + perception policyBuilt and optimized by Haink

~$15,000–$45,000 compute + perception + software per node, ex-works hub; excludes the mobile base

VLA Inspection Cell

Visual quality inspectionMulti-cameraJetson AGX ThorReasoning VLM

A high-end perception cell running a vision-language-action or reasoning vision-language model (such as NVIDIA Cosmos Reason) on Thor-class compute, fusing several cameras to detect defects and drive a reject or sorting actuator — for tasks too varied for a fixed vision program.

ComponentSpecificationNotes
Edge computeNVIDIA Jetson AGX Thor128 GB; runs VLA / reasoning VLM
PerceptionMulti-camera array (+ depth)Synchronized real-time fusion
ActuatorReject / sort mechanism or armActs on the model’s decision
NetworkingHigh-bandwidth (RDMA where needed)For multi-sensor streams
SoftwareVLA / reasoning VLMBuilt and optimized by Haink

~$50,000–$150,000 delivered cell, ex-works hub; Jetson AGX Thor devkit alone ~$3,499

Digital-Twin → Edge Development Rig

Sim-firstRTX 6000 AdaTrain → deploy

The bench to build a deployment before metal moves: a GPU workstation for the digital twin and training, plus one edge node as the deployment target — so the model is validated in simulation, then transferred to the real machine. Pairs with digital twin and training.

ComponentSpecificationNotes
Simulation / trainingRTX 6000 Ada-class workstationTwin + synthetic data + training
Deployment targetJetson AGX Orin / ThorValidate transfer on real hardware
SoftwareOmniverse / Isaac Sim + training pipelineBuilt by Haink

~$15,000–$30,000 ex-works hub; RTX 6000 Ada workstation ~$12K + edge node (Thor devkit ~$3,499)

Need hardware only? Browse robotics & physical-AI hardware →  ·  Want it built end-to-end? See robotics deployment →

How to read these

Each blueprint lists the compute, perception, actuation and software for one task pattern, with an indicative "from" price. They are sized to a typical case — your reach, payload, cycle time, sensor count and safety requirements move the BOM and the price. Send your task and we return a tailored bill of materials, firm pricing and compliant lead times, screened for export compliance per destination.

Frequently asked questions

What is a physical AI reference architecture?

A validated blueprint for a robotics deployment — the edge compute, sensors, actuator and software for a specific task such as a pick-and-place cell, an autonomous mobile robot or a VLA inspection cell — with a bill of materials and indicative pricing you can budget from.

How much does a robotics cell cost?

Indicatively, a delivered edge + manipulator cell runs ~$40,000–$150,000 (the cobot arm alone is ~$25K–$60K in 2026), an AMR compute-and-perception node ~$15,000–$45,000 (excluding the mobile base), and a VLA inspection cell ~$50,000–$150,000 — depending on the robot, sensors and compute chosen. See also our deployment cost guide.

Can you tailor a reference architecture to my task?

Yes. Each blueprint is a starting point; Haink sizes the compute, sensors and actuator to the task and returns a tailored bill of materials with firm pricing and lead times, then builds the model and supplies the hardware.

Do these include the AI model or just the hardware?

Both. Haink builds and optimizes the perception, control or vision-language-action model and supplies the edge and robot hardware it runs on, as one accountable partner.

Need this sized to your task?

Send the task, environment and constraints — we return a tailored BOM, firm pricing and compliant lead times.

sales@haink.org