IN Brief:
- The 22 October technical disclosure follows the initial Iris 1.0 announcement on 15 October.
- Iris-1 uses a video transformer adapted from Meta’s V-JEPA architecture and supports operational event searches.
- OneTrack reports 90% collision detection against its selected test set; the benchmark was not independently reproduced.
OneTrack has published a technical account of Iris-1, its specialised artificial intelligence model for interpreting video from warehouses and industrial operations. The system identifies events involving material handling equipment and allows users to search recorded activity through natural language, supporting investigations undertaken through the company’s existing monitoring software.
The disclosure dated 22 October follows an initial Iris 1.0 announcement on 15 October. The second publication concentrates on the model’s architecture, processing costs and testing methodology, adding technical information to a product that had already been introduced rather than presenting a separate first launch.
Across hundreds of facilities, the company says its installed systems process more than 150,000 operational events daily. The footage provides examples of the movement and interaction of forklifts, loads and other equipment in industrial settings, creating training material more specialised than general-purpose image datasets.
Recognising an incident requires analysis of how a scene changes over time. A forklift may pass close to a storage rack without making contact, while a sudden change in movement could indicate either a collision or an ordinary manoeuvre. Interpreting the sequence of frames is necessary before classifying the activity.
Iris-1 uses an architecture derived from Meta’s V-JEPA 2.1 video model, with a transformer containing approximately two billion parameters. OneTrack says it has adapted and trained the system using proprietary anonymised industrial footage to improve its representation of physical activity in warehouses.
The encoder processes a recorded event into an internal representation that can be used when questions are asked about what happened. A separate component links that representation with language queries, allowing users to request descriptions or locate footage involving selected events without repeatedly encoding the original recording for every question.
This approach can reduce repeated computation when several queries concern the same event, although the effective saving depends on hardware, workloads and the number of questions asked. More importantly, the answers remain constrained by what was visible in the footage and whether the model correctly interpreted the recorded movement.
The model also uses a confidence mechanism that compares a proposed interpretation against similar events reviewed previously. Where the available examples offer weak support, the system can indicate uncertainty. Such a mechanism depends on the breadth and quality of the reference material and does not eliminate the need to investigate ambiguous incidents.
To evaluate the model, the company tested Iris-1 against selected general-purpose AI models using industrial footage excluded from training. On a collision recognition task, OneTrack reports detecting 90% of genuine forklift collisions, compared with 29% for Google Gemini 3.8 Flash and 7% for Anthropic Claude Sonnet 5.5 under its chosen test conditions.
Those percentages describe recall on the selected collision examples, not overall performance across every safety event. A model might correctly identify an impact while misidentifying the object struck, and its description may contain both accurate and inaccurate details. OneTrack’s additional reported results include 74% correct identification of the object involved in a collision and 75% factual accuracy in an open description task.
The benchmark was designed and administered by the company and has not been independently reproduced across a representative range of facilities. Camera placement, lighting, obstructions and differences between equipment types can all influence whether an event is visible and correctly recognised. The published results therefore require qualification when considering performance in a particular warehouse.
Processing cost is another element of the disclosure. OneTrack estimates approximately $0.00003 per answer based on graphics processing hardware electricity consumption, sharing the initial encoding cost across queries on the same event. The comparison omits hardware purchase, staffing and data labelling costs, and does not represent the total price of an operational monitoring deployment.
Its cost comparison further illustrates a hypothetical warehouse with 100 forklifts and a specified daily event count to estimate potential annual processing costs. Those calculations depend on assumed workloads rather than audited expenditure, while the actual cost for a customer would include the infrastructure and services required to collect, store and analyse video.
Within OneTrack’s platform, Iris-1 supports searches for activity involving forklift impacts, dropped loads and other operational incidents. Its AiOn agent can connect information extracted from recordings with wider event data, making footage easier to locate during a safety review or an investigation spanning several facilities.
A video record can show the movement that preceded an impact without establishing why it occurred. Equipment condition, route layout, operator instructions and surrounding traffic may need separate examination before an organisation can determine the underlying cause of an event. The model’s classifications are therefore investigative inputs rather than conclusive findings.
Possible future applications include yards, ports and other industrial settings. Those intentions are separate from the currently available functions tied to its existing monitoring platform. The October publications establish the system’s design and selected internal results, with broader operating performance still dependent on deployment conditions and independent evaluation.
