How-to

How to check missing assembly components with two cameras and a reviewed evidence trail

The short answer

A missing-component check should compare a named assembly recipe with visible image regions and preserve the evidence behind each result. Two cameras can cover separate views, but cannot guarantee visibility of hidden or occluded parts. This proposed design keeps the checker on customer-operated edge hardware and sends only approved evidence to HyperLake in the customer's cloud account.

How should the expected assembly be defined?

Begin with the bill of materials and the station's inspection responsibility. Define exactly which component is visible from each viewpoint, which variants are allowed and what should happen when the part is obscured. Do not use one generic “complete” label across changing products.

Bind the inspection to a recipe identifier supplied by an approved production source. A stale recipe can produce confident errors even when the model works as designed. Log recipe version, fixture identifier, camera calibration, model version and operator disposition for every unit.

Recipe identity, view coverage and uncertainty handling stay attached to the inspection record.

How are classification and view coverage different?

An image classifier can be one part of the check. Torchvision documents ResNet-18 with 11,689,512 parameters and 1.81 GFLOPS for its documented configuration. These describe that reference model, not assembly-inspection accuracy or throughput. A task-specific model needs relevant examples, evaluation and preprocessing that preserves the evidence.

Use a deterministic alignment check before interpreting component crops. If the fixture is misaligned, mark the inspection incomplete. Count expected views and crops so that missing input cannot accidentally become a pass. Add an independent check for critical parts that cannot be verified visually.

What do the two-camera volume assumptions imply?

Illustrative estimate: assume two cameras at 1920 × 1200 pixels, one RGB8 frame per view per assembly, two assemblies per second, eight operating hours and four component regions per camera. The camera resolution is listed for the Basler acA1920-40gc; counts, rates and storage format are scenario assumptions.

Assemblies per day = 2 × 8 × 3,600 = 57,600. Two views create 115,200 frames. With 6,912,000 uncompressed bytes per RGB frame, retaining every view uses 796.2624 GB per day, excluding metadata and replicas.

Eight regions per assembly at two assemblies per second create 16 crop evaluations per second. Multiplying that rate by the published 1.81-GFLOPS ResNet-18 reference gives 16 × 1.81 = 28.96 GFLOPS of model operations per second in this simplified arithmetic. It omits alignment, decoding, crop extraction, transfer, runtime overhead and differences in the actual model. Do not convert it into a GPU utilization or cameras-per-GPU promise.

If a policy retains one full two-view record for 2% of assemblies, the image subtotal is 15.925248 GB per day. Two percent is a retention assumption, not a defect rate. Uncertain results, changing product mix and traceability rules may require more retention.

The arrival interval is 500 milliseconds per assembly. Measure the local decision deadline and test queue recovery after a burst or interruption. The product arrival rate alone does not establish how much time remains before a human or actuator needs the result.

What is a sample compute cost for the inspection cell?

Illustrative compute-module estimate, not a complete inference cost or quote: assume one NVIDIA Jetson Orin NX 8 GB module per camera, two cameras per cell, a three-year straight-line hardware allocation, a 40 W module power mode, eight operating hours per day and 250 operating days per year. NVIDIA lists a $649 USD volume suggested price at 1,000 units and documents the 40 W mode for Orin NX production modules. That is the module price, not a complete edge system. EIA reports a 2025 U.S. industrial electricity average of 8.51 cents per kWh; apply the facility's actual tariff for a local estimate.

Per camera, the annualized module allocation is $649 ÷ 3 = $216.33. Assumed module electricity is 0.040 kW × 8 hours/day × 250 days/year = 80 kWh, or 80 × $0.0851 = $6.81 per camera-year. For two modules, the subtotal is $446.28 per cell-year ($223.14 per camera-year). It excludes cameras, lenses, lights, carrier boards, enclosure, power conversion, the rest of the edge computer, installation, cooling, cloud, storage, software, support and tax. Cost arithmetic does not show that the modules can run the assumed 16 crop evaluations per second; model fit and end-to-end capacity remain unmeasured.

For an annual cell estimate, use cell count × (quoted complete two-camera edge-computer cost divided by its chosen service life + measured annual computer kWh × the site's tariff), then annualize quoted one-time cameras, optics, lighting and installation over their chosen service lives; add annual recurring labor, software/support, cloud/storage/egress and tax as applicable. Use the module-only floor above only when no complete-computer quote exists; do not add it again if that quote already includes the modules. No cost per inference is claimed without measured throughput.

What should a pilot include and test?

The proposed cell needs two cameras with compatible triggers, two lenses, controlled lighting, a stable fixture, assembly and recipe identifiers, an edge computer with buffer storage, and an operator interface. Cloud-side evidence storage is optional until the permitted upload scope is agreed. Select accelerator memory and runtime after measuring the trained model and image pipeline.

Acceptance tests should include a correct assembly, each deliberately missing visible part, allowed variants, occlusion, shifted fixtures, empty frames, duplicated triggers and an incorrect recipe. Review false accepts separately from false rejects. Give the quality owner authority to reject a model change even when average metrics improve.

For each test, retain the expected recipe, the source images, the crop boundaries and the reviewer's decision. Group challenge examples by variant and capture condition so a strong result on one common assembly does not conceal a miss on a less common option. Keep a “not inspectable” result distinct from “component absent.” If reviewers disagree about whether a part is visible or acceptable, resolve the label rule before tuning a threshold. Start with the checker in observation mode and compare its suggestions with the established station procedure before it affects the operator's disposition.

How should recipe governance and safety be handled?

Recipe editing is a controlled change. Separate permission to view results from permission to change expected components or acceptance thresholds. Protect audit records from casual alteration and test rollback to a known-good recipe and model.

This proposal does not establish functional-safety or product-release certification. In the US, OSHA's general machine-guarding rule addresses protecting operators and other employees; this image checker is not a guard or safety device. In Canada, identify the applicable provincial or federal workplace-safety regime for the site. Keep safety interlocks and personnel protection outside an unevaluated vision workflow. The NIST OT security guide can inform security review when an inspection application connects to production systems.

Where views include employees, badges or product serial numbers linked to people, minimize unnecessary capture and define access and retention. Canada's privacy commissioner says PIPEDA applies to commercial activities and to employee personal information in federally regulated businesses; determine its scope for the organization and review relevant provincial rules. The commissioner's employee-video case concerns an employer subject to PIPEDA, not every Canadian employer. In the US, review the site's employee-video and state privacy requirements. Local inference or cloud residency alone does not establish compliance.

Where does HyperLake fit, and when does another route fit better?

HyperLake runs in the customer's own AWS or Google Cloud account today. Azure, OVH, private cloud and on-premises deployments are coming soon. Its documented scope can support an agreed cloud-side evidence, data and model environment; this page does not claim a prebuilt production-line connector or validated component-detection model.

This approach may suit teams that need consistent evidence, identity and controlled model operations across multiple cells. A specialist inspection system may be preferable when it already supports the required components and line integration. A currently available local-only stack is a better fit when all evidence and operation must remain on site.

How can a first pilot stay bounded?

Run a 2-week pilot with one approved recipe, an independent challenge set and a defined operator fallback. This is a proposed trial period, not a delivery guarantee. See the physical-AI hub, the factory inspection architecture and the companion packaging surface-defect review. The LLM calculator applies only to a separate language-model feature such as drafting a review summary; it does not estimate component-detection cost.

Frequently asked questions

Do two cameras guarantee every component is visible?

No. Coverage depends on placement, geometry and occlusion. Document which components each view can establish and use another method for hidden components.

Can a pretrained image classifier verify an assembly immediately?

No task-specific accuracy is established here. The model, preprocessing, recipe and acceptance rules need training or adaptation and independent evaluation.

Is the operations estimate a GPU-sizing result?

No. It multiplies a published reference-model operations figure by an assumed crop rate and excludes much of the real pipeline. Size hardware from measured end-to-end behavior.

What happens if the wrong recipe is loaded?

The system should expose the mismatch or stop the inspection according to the agreed procedure. Record the recipe source and version; do not silently use the last recipe.

Does this require uploading every assembly image?

No. The proposed upload policy can retain selected evidence, but required traceability may call for more. Agree the retention and audit design before using a reduced sample.

Sources

Physical AI infrastructureWhat a Physical AI stack needs, why sovereignty matters for robot data and models, and how Hyperlake deploys and governs the stack in your environment.
Data infrastructure for physical AI and robotics
Robotics training-data formats and provenance

Start with a workload. Build the environment around it.

What do you need to deploy? Whose environment does it have to run in, and what does it connect to? Tell us.

Discuss a deployment ↗Explore use cases