Bucket Robotics wants to fix industrial vision’s worst bottleneck: the data

Bucket Robotics came to CES 2026 with a focused pitch: train surface-defect inspection models from CAD files instead of spending weeks collecting and labeling real production images.

That may sound narrow. It’s not. In factory vision, the slow part is usually the dataset. Cameras are easy to buy. Integrating with a line is painful but familiar. Building labeled examples for every scratch, scuff, burn mark, color mismatch, and raised blemish across every part revision and finish is where projects stall.

Bucket’s answer is synthetic data generated from the part’s own geometry. If that holds up on real production lines, it’s a credible way into automotive, defense, and other manufacturing environments where visual inspection still depends too much on human judgment.

Why the timing works

Manufacturers are under pressure to add capacity closer to home. They’re doing it with tighter labor markets and less patience for long integration cycles. Surface inspection has stayed stubbornly manual because cosmetic defects are messy. A dimensional check is straightforward. A visual defect depends on lighting, material, viewing angle, severity, and whether the customer will actually reject it.

That’s the problem Bucket is chasing.

Founder Matt Puchalski comes from autonomous vehicles, with stints at Uber, Argo AI, Ford’s Latitude AI, and Stack AV. That background matters. AV teams spend years dealing with sensor modeling, synthetic data, calibration, and sim-to-real transfer. Moving that playbook into factory vision makes sense, especially for a company selling deployment speed instead of a broad robotics platform.

At CES, that drew interest from automotive and defense buyers. No surprise there. Those industries care about quality escapes, traceability, and frequent line changes. They also know how painful custom vision pipelines can get.

Why CAD matters

Using CAD as the starting point gives the system something solid to work from.

A part’s CAD model contains the nominal geometry, tolerances, surface boundaries, and often some material context. That gives an inspection stack a map of where defects should appear and how a camera should see them once the image is aligned to the part pose. From there, you can synthesize cosmetic damage directly on the model, render it under different conditions, and train a supervised model with labels you didn’t have to create by hand.

That’s the attraction. No team of annotators drawing boxes around tiny scratches. No waiting for enough real defects to show up in production before training can even start.

For a line engineer, the pitch is straightforward: send the CAD, connect the system to the existing camera feed, and get to a usable detector quickly.

If Bucket can do that consistently, it matters.

The synthetic data has to be honest

This is where industrial vision projects usually run into reality.

Generating fake defects from CAD is easy to describe and hard to do well. Surface inspection lives in bad lighting, reflective materials, brushed metal, slight texture variation, shadows from fixtures, and the accumulated sins of factory optics. A model trained on clean synthetic images will fall apart quickly if it lands on a real line full of glare, dust, aging LEDs, and parts that never sit quite the same way twice.

So the rendering stack matters. A serious pipeline probably needs:

• Physically based rendering for materials and lighting
• Local mesh or texture perturbations for scratches, chips, bumps, and scuffs
• Domain randomization across camera pose, exposure, noise, and background clutter
• Some kind of photometric calibration or limited fine-tuning on real images

That last point is where any “minutes to deploy” claim needs scrutiny. Synthetic labels from CAD can cut setup time. They probably don’t eliminate calibration, validation, and threshold tuning in a production setting where reject rates cost real money.

Factories don’t care that the benchmark looks good. They care whether the system starts rejecting good parts at 2 a.m.

Harder than anomaly detection

A lot of industrial AI vendors have leaned on anomaly detection. Train on normal parts, then flag anything that looks off. Tools and papers like PaDiM, PatchCore, and CFlow can work well when defects are rare and labels are scarce.

But anomaly detection only gets you so far in regulated manufacturing. Quality teams usually want an explicit defect class, a severity score, and a threshold tied to an engineering spec. Is this a burn mark? How long is the scratch? Is the blemish acceptable for this finish grade? Can someone audit the decision later?

That’s where a synthetic supervised approach is more compelling. If you can train on labeled simulated defects tied to specific failure modes, the output is easier to explain and easier to plug into existing quality workflows.

It also fits the rest of the plant stack. Defect classes and severity thresholds can feed into SPC, MES, or quality gates on the line. Results can be stored with part ID, revision, and finish code. In automotive and defense, that kind of traceability is mandatory.

The hardware claim matters too

Bucket says customers don’t need new cameras or lighting rigs. If that holds up in a decent number of cases, it may be the strongest part of the pitch.

Industrial vision projects often fail in integration, not in model training. New camera systems bring downtime, electrical work, validation, operator retraining, and plenty of internal pushback. If a vendor can use existing feeds and still get useful results, pilots get approved faster.

There’s an obvious limit. Some surfaces are awful. Highly reflective finishes, anisotropic brushed metal, and subtle texture defects often need specialized imaging setups such as polarized light, photometric stereo, or deflectometry. If Bucket pushes the idea that no new hardware is ever needed, that would be hard to take seriously. Physics still applies.

Still, a lot of plants aren’t trying to solve the hardest optics problem on day one. They want acceptable inspection for common cosmetic defects without a six-month retrofit. For that, working with installed cameras is a strong selling point.

What technical teams should watch

The interesting part isn’t the phrase “synthetic data works.” It’s the operational stack around it.

A usable system here needs tight control over alignment, latency, and model governance.

Pose estimation and calibration

CAD-driven inspection only works if the image aligns to the part well enough to define repeatable regions of interest. If fixtures drift or camera position shifts, defect localization degrades fast. Pose estimation against CAD, plus regular camera calibration, is basic plumbing.

Fast retraining for part revisions

This is one place Bucket could stand out. Manufacturing lines change constantly. New revisions, new finishes, small geometry changes. If the synthetic pipeline can regenerate data and retrain in hours instead of weeks, that’s real value. It makes model maintenance feel closer to a build pipeline than a custom vision project.

Edge inference

Inspection has to run at line speed. That usually means compact segmentation or detection models, quantization, and deployment on edge hardware like Jetson-class systems or optimized x86 boxes running TensorRT, OpenVINO, or similar stacks. Average latency isn’t enough. Worst-case latency matters, because conveyors and reject mechanisms won’t wait for p95.

Security and IP

CAD files are sensitive. In some sectors, they’re among the most valuable files in the building. Any vendor building on CAD-native training has to answer the obvious questions around storage, access control, retention, and whether models can leak geometry or proprietary process details. Defense customers will care a lot. Everyone else should care too.

A crowded market, but a smart angle

Bucket isn’t walking into an empty category. Companies like Landing AI, Instrumental, Scortex, and Covision Quality have all pushed machine learning for industrial inspection in different ways. What stands out here is the target. Bucket is going after the part customers hate most: data prep for every new variant.

That’s a smart place to start. In industrial AI, “faster deployment” often means faster than an already miserable baseline. A CAD-first synthetic workflow could actually compress one of the longest parts of the cycle.

It also gives customers a cleaner maintenance path. If the source of truth is the engineering model instead of an image dataset collected months ago, updates are easier to reason about. New part revision? Update the CAD, regenerate defects, retrain, validate, ship.

The company still has to prove it under ugly real-world conditions. Reflective parts, inconsistent plant lighting, mixed suppliers, old cameras, and edge-case cosmetic defects are where a lot of confidence disappears.

Worth watching for a reason

There’s no shortage of AI companies trying to modernize manufacturing. Many underestimate how stubborn factory conditions are.

Bucket’s pitch lands because it starts with a real pain point and a plausible shortcut. CAD already exists. Synthetic rendering has improved enough to be useful. Sim-to-real workflows are better understood than they were a few years ago. Buyers are also more willing to try software that fits into existing lines instead of forcing a hardware overhaul.

None of that guarantees success. Surface inspection is still one of the harder vision problems in industry, and a clean demo doesn’t tell you much about three shifts, plant drift, and supplier changes.

Still, this is the kind of narrow technical wedge that can turn into a real business. If Bucket can show low false-positive rates, fast model updates, and solid traceability on existing production hardware, it won’t need CES buzz for long. Quality engineers will have a reason to pick up the phone.

What to watch

The caveat is that agent-style workflows still depend on permission design, evaluation, fallback paths, and human review. A demo can look autonomous while the production version still needs tight boundaries, logging, and clear ownership when the system gets something wrong.