01.Summary
Encord has extended its collaboration with NVIDIA to bring synthetic data generation to Physical AI training workflows.
Models are only as reliable as the scenarios they have seen, and the scenarios that matter most are the ones rarely captured. Synthetic data closes that gap, enabling teams in any industry to generate rare scenarios from images they already have and calibrate detection before a failure occurs. Working with ANYbotics, a leader in autonomous inspection for oil and gas, energy and manufacturing sites, Encord deployed newly launched NVIDIA Event Video Generation (EVG) skill on industrial production imagery ahead of its launch at NVIDIA GTC Berlin 2026.
In our testing, synthetic anomalies improved recall for objects appearing by 24% and boosted detection accuracy for liquid spills by 10%. ANYbotics estimates the gain on spills alone to be worth over $22,000 per site annually, more than 400 times the cost of generating the data itself. The work builds on existing NVIDIA Cosmos and Cosmos Embed integrations in Encord for pre-labeling and behavior search: together, they give Physical AI teams one platform to find gaps in their data, generate what’s missing and enrich the results, all with the latest NVIDIA models.
02.The challenge
ANYbotics’ ANYmal fleet is deployed at sites worldwide in environments too hazardous or remote for routine human patrols. On each mission, the robots return to the same inspection points looking for anomalies: rare changes in a scene that lead to downtime, damage or safety incidents if they go unseen. ANYbotics’ change detection model compares each capture against the distribution of normal images for that point, flagging any that fall outside it. A well-run site produces very few anomalies, and an ANYmal can patrol for months without one, so distribution boundaries are often set without a real example to calibrate on: by the time one exists, the failure has already happened. Leveraging synthetic anomalies generated from each point’s own images allows ANYbotics to analyze how anomalies differ from normal prior to encountering them in the real world.
03.How it works
Part of NVIDIA Cosmos Data Factory and powered by NVIDIA Cosmos, the EVG skill turns a single reference image into a short clip with a requested change. A vision-language model captions each reference image before it is combined with the change type assigned to the inspection point, forming the prompt Cosmos generates from. Multiple change types can be generated from a single scene. In our work with ANYbotics, EVG produced over 1,500 clips from 410 reference images across nine change types. ANYbotics took multiple frames from each clip to calibrate its change detection model, giving several samples and degrees of change per generation.
With Encord, teams can curate reference imagery before generation (grouping by inspection points and their associated change types) and verify auto-labeled outputs before training. Annotations generated by Cosmos Data Factory map to existing ontologies in Encord, and there is full provenance across the entire pipeline: every generated clip carries metadata including the requested change, the prompt used and its automated quality verdict. Teams can audit synthetic data to the same standard as real data, with nothing ever leaving their own cloud.
NVIDIA Cosmos Data Factory + Encord
01Curate
Reference images grouped by inspection point with their change types
Encord
02Prompt
A caption of the scene plus the requested change type
NVIDIA Cosmos Data Factory
03Generate
Event Video Generation, powered by NVIDIA Cosmos
NVIDIA Cosmos Data Factory
04Verify and label
Quality scoring; labels mapped to the customer's ontology
Encord
05Deliver
Before-and-after pairs, inside the customer's own cloud
Encord
Every clip carries
04.Results
Our evaluation tested the performance of synthetic data against real anomalies, comparing distribution boundaries set from normal images with those also including generated frames. Across all inspection points that received synthetic data, anomaly recall rose 24%. Liquid spills and objects appearing showed the largest gains, with detection accuracy increasing by 10% for both change types. Recall increased by 8% and 24%, respectively, and example volume for objects appearing rose by 221%.
Recall on real anomalies, with and without synthetic data
- Normal images only
- With synthetic clips
| Category | Normal images only | With synthetic clips |
|---|---|---|
| Objects appearing, 23 inspection points, +24% | 0.54 | 0.67 |
| Liquid spills, 9 inspection points, +8% | 0.47 | 0.50 |
| All classes, 25 inspection points, +24% | 0.50 | 0.62 |
Anomaly examples available for calibration, real and synthetic
- Real anomalies
- Synthetic clips
| Category | Real anomalies | Synthetic clips |
|---|---|---|
| Objects appearing, +221% | 237 | 525 |
| Liquid spills, +45% | 205 | 92 |
ANYbotics assessed that real anomalies remain stronger calibration data where they exist, but since they are sparse and unknown in advance, synthetic data gives a meaningful head start on change types with few or no real examples. For industrial use cases, this is clearest with liquid spills and objects appearing, which are less common in real world data but simple to verify. Calibrating on synthetic anomalies also raised the false alarm rate on normal scenes, with some change types not meaningfully improving; the team hypothesized that using normal augmentations (changing conditions such as lighting, angles and blur with the NVIDIA Video Data Augmentation skill) alongside generated anomalies could help close that gap by preventing ordinary changes from reading as anomalies.
Return on generation was also high, making synthetic data a cost-effective way of de-risking industrial sites. The data used to increase liquid spill recall by 8% cost just $55 to generate, around 60 cents per clip. ANYbotics values that gain at over $22,000 per site annually — over 400 times the cost of generating it with the EVG skill. At higher risk sites the value climbs, reaching $31,000 at a hydroelectric power station and $78,000 at an offshore oil and gas facility, through avoided equipment damage, downtime, safety incidents and environmental fines.
ROI on liquid spill generation
- Cost to generate the clips, once
- Value of catching 8% more spills, per site per year
| Category | Cost to generate the clips, once | Value of catching 8% more spills, per site per year |
|---|---|---|
| Generation cost | $55 | No data |
| Chemicals plant | No data | $22,800, 415× |
| Hydro power station | No data | $31,400, 571× |
| Offshore oil and gas | No data | $78,500, 1,427× |
05.What’s next
This approach generalizes beyond industrial inspection, extending to workplace safety, manufacturing, warehouse operations, utilities and any other setting where the scenarios that matter most are the rarest. Any vision system that needs to detect such scenarios can generate examples from the scenes it already captures with Cosmos Data Factory, then verify, label and enrich them alongside real data in Encord. The pipeline helps teams calibrate models ahead of a failure occurring, for a small fraction of what a single missed failure costs — all without the underlying data leaving a customer’s own cloud.
Looking ahead, Encord and NVIDIA are working on tighter integration of Cosmos Data Factory into the Encord platform, with generation sitting alongside curation, auto-labeling and review.
Subscribe for new releases
Research at the frontier.