OneTrack Introduces Iris 1.0, a Domain-Specific Vision Model for Industrial Environments
Industrial operations record far more than anyone can watch. Iris 1.0 understands every event in context, describes it in plain language, and gives AI agents a reason to act. Version 1.0 starts in the warehouse.
OneTrack.AI, a leader in agentic AI for physical operations, today introduced Iris 1.0, a domain-specific model for frontier vision intelligence in industrial environments, starting in the warehouse. Trained on one of the largest data sets of real-world operational data, hand-curated, selected, anonymized and cleaned, Iris 1.0 reads each event the way an experienced operator would. It is available today in the OneTrack dashboard and in AiOn, OneTrack's agentic platform for physical operations.
Why industrial environments need their own model
OneTrack sensors record more than 150,000 events a day across hundreds of warehouses, far more than any team can watch. Understanding them takes context: motion, direction, and contact, seen from above. General-purpose vision models, trained mostly on phone and social video, describe the scene. Iris 1.0 describes what the lift did.
On OneTrack's own warehouse benchmarks, Iris 1.0 averaged 73% balanced accuracy across four safety questions, compared with 41% and 50% for two leading general-purpose models, at a fraction of the cost per answer.
General-purpose models describe the camera. Iris 1.0 describes what the lift did. That comes from context, and three years of footage from the warehouse floor is context nobody else has.
What Iris 1.0 does today
- Describes every event in plain language, searchable in plain English.
- Gives AI agents a reason to act. Agents check every new event and act when something matters, without waiting to be asked.
- Runs on every event for about $0.00003 per answer.
Most of what happens in industrial operations has always gone unseen. Iris 1.0 changes that, and because it's our own model, it keeps learning with every label our team adds.
Iris 1.0 is available now for events from overhead sensors, with more camera views and industrial environments to follow.
Key facts
- What
- OneTrack.AI introduces Iris 1.0, a domain-specific model for frontier vision intelligence in industrial environments. Version 1.0 starts in the warehouse
- Training
- One of the largest data sets of real-world operational data, hand-curated, selected, anonymized and cleaned, plus 505,000 descriptions grounded in each sensor's accelerometer and gyroscope data
- Model
- A 2-billion-parameter vision transformer with a lightweight predictor that answers directly in language space, with no large language model in the loop
- Scale
- OneTrack sensors record more than 150,000 impact events a day across hundreds of warehouses
- Benchmarks
- 73% balanced accuracy averaged across four warehouse safety questions, compared with 41% and 50% for two leading general-purpose vision models
- Cost
- About $0.00003 per answer on OneTrack's own GPUs, more than 60 to more than 1,400 times less per answer than the general-purpose vision models OneTrack tested
- Availability
- Available today in the OneTrack dashboard and in AiOn, OneTrack's agentic platform, for events from overhead sensors
- Date
- Oct. 15, 2026
Frequently asked questions
What is Iris 1.0?
Iris 1.0 is OneTrack's domain-specific model for frontier vision intelligence in industrial environments, starting with the warehouse floor. It watches the clip from each forklift event captured by OneTrack sensors and describes what happened in plain language, so every event can be searched, reviewed, and acted on by people and by AI agents.
What does a domain-specific vision model mean?
A domain-specific vision model, sometimes called a world model, learns how one kind of environment works from large amounts of data about it. Iris 1.0 learned industrial work starting with the warehouse, from one of the largest data sets of real-world operational data, hand-curated, selected, anonymized and cleaned, seen through OneTrack's overhead fisheye sensors, and from motion data measured by the sensor itself, rather than from general web and social video.
Why does version 1.0 start in the warehouse?
The warehouse is where OneTrack's data is deepest: three years of footage from sensors across hundreds of warehouses. Iris 1.0 works on events from overhead forklift sensors today, with more camera views and industrial environments to follow.
How is Iris 1.0 different from general-purpose vision AI?
General-purpose vision-language models describe the scene, and often the camera. Iris 1.0 describes what the lift did: the turn, the reverse, the contact. It answers directly in language space with no large language model in the loop, runs on OneTrack's own GPUs, and shows through heat maps which part of each frame an answer came from.
How much does it cost to run?
Iris 1.0 runs at about $0.00003 per answer, the electricity cost of its GPU time. In OneTrack's internal testing, the general-purpose vision models it compared cost more than 60 to more than 1,400 times as much per answer, which is what makes it practical to describe every event, every day.
Where can customers use Iris 1.0?
Iris 1.0 is available today in the OneTrack dashboard, where every event comes with a description and can be searched in plain English, and in AiOn, OneTrack's agentic platform, where AI agents watch for specific behaviors and act when they occur. It currently works on events from overhead sensors.
About OneTrack.AI
OneTrack builds the agentic AI operating system for the physical world: software written by agents, used by humans, and trusted at scale. Its sensors capture ground truth on the warehouse floor, and its AI agents connect that data to the systems that run an operation, so teams can build, ask, and automate in plain language, securely and at scale. Learn more at onetrack.ai.
Media contacts
Kira Gyöngyösi, Director of Growth, OneTrack.AI, media@onetrack.ai