Originally published 2026-11-13 · VWU.AI
Physical AI in Korea: Building a Traceable Robot Video Data Loop
Connect observed failures, follow-up data, annotation versions and independent testing in a robot video workflow.

NVIDIA’s November 8, 2026 report describes an LG CLOiD manufacturing demonstration and a workflow linking real capture, Cosmos augmentation, GR00T training, Jetson compute and simulation evaluation. This is a vendor account. The workflow below is an engineering proposal and does not infer capability from the demonstration.
Define the observed failure
Specify what the robot was asked to do, what outcome was observable, and how the event was judged. Separate task failure, missing evidence, annotation errors and hardware or control problems.
Create a failure record
Link the original recording, time interval, task and versions, camera and visibility, applicable rules, reviewer decision and proposed follow-up. Let capture staff find what to collect, annotation staff see how to judge it, and evaluators reproduce the comparison.
Separate intent, execution and outcome
Record the command, observable stages such as approach or contact, and final outcome independently. A commanded placement is not evidence that the item reached the bin. Keep negative and indeterminate events, and verify synchronization between cameras and action logs.
Investigate a hypothetical slipping container
If a container slips during transport, review whether tilt, phase or occlusion is relevant before collecting new colors. Keep the original condition as a real reference and vary one relevant factor at a time in controlled testing. Add a view or verifiable record if contact cannot be observed.
Connect data versions to a controlled comparison
Record the baseline version, the new samples, training settings and test selection. Compare real follow-up and synthetic expansion with a fixed baseline where appropriate. Keep related objects, scenes and adjacent footage together in splits and hold out real test cases from data and model tuning.
Use independent evidence to close the loop
Related generators and evaluators can screen samples, but real independent tests should establish performance. A feedback round may lead to revised labels, camera placement, better logs or hardware changes. Start with one failure that another engineer can trace and reproduce.