Vision Lab Customers | AI Video Labeling for Robotics Teams

From our first cohort

Examples from early-access teams using Vision Lab on their first projects.

Warehouse robot obstacle detection data

A warehouse robotics team needed labeled video for a mobile robot obstacle detection model. Four part-time annotators had been handling the footage manually, while a 3 weeks backlog was holding up model iteration and training.

They uploaded the footage to Vision Lab, then set six categories: pallets, shelving units, people, forklifts, and emergency stop zones. The AI pass ran overnight and sent 8% of frames for human review. The team cleared that queue in 2 days, then exported a COCO-formatted dataset with a coverage audit across all six categories.

Backlog cleared in 5 days, versus a projected 3 weeks

Warehouse automation team, early-access cohort, Q1 2026

Agricultural harvesting arm data

An agricultural robotics startup needed segmentation masks, not bounding boxes, for an autonomous harvesting arm. The team also had limited annotation capacity while engineering work continued.

They configured Vision Lab for segmentation masks and defined classes for produce types, stem geometry, and occluded objects. A coverage report showed that one produce variant was 3x below its target density. They collected more targeted footage before the second pass, catching the gap before training rather than after it.

Coverage gap found before training, saving an estimated 2-week cycle

Agricultural robotics startup, early-access cohort, Q2 2026

Industrial robot inspection quality control

A robotics team building an automated inspection system for a manufacturing client needed video data for defect detection. Its taxonomy contained 14 defect categories, with severity distinctions that required reviewers to apply the labels consistently.

Vision Lab's inter-annotator consistency scoring revealed disagreement in 2 categories during review. The team revised the taxonomy, then reviewed the affected data again before export. This addressed annotation drift before it could affect model behavior.

Inter-annotator agreement rose from 0.71 to 0.93 after taxonomy refinement

Industrial inspection robotics team, early-access cohort, Q2 2026

From early-access teams

We had three weeks of footage waiting because annotation kept blocking us. Vision Lab cleared it in 5 days. The final coverage report was especially useful: it showed we needed more emergency stop zone footage before training, instead of finding that gap six weeks later when the model failed.

R

Robotics Lead

Warehouse automation team, early-access cohort

Manual segmentation takes too long. The AI pass handled the clear cases and sent ambiguous ones to review. We spent 2 days on focused review, not 3 weeks labeling everything.

ML

ML Engineer

Agricultural robotics startup, early-access cohort

The consistency score found an issue we might have shipped. Two reviewers were using different bounding boxes for the same defect class. Finding that during annotation costs far less than finding it after model evaluation.

EI

Computer Vision Engineer

Industrial inspection team, early-access cohort

Early Access

Apply for early access

We review each request against the platform's current capabilities. If the use case fits, we will contact the team to discuss a practical starting point.