IN Brief:
- Dexory has incorporated NVIDIA Cosmos-Reason2-2B into DexoryView Storage Health.
- The model provides prompt-guided analysis when existing image classifications have low confidence.
- The 2B version was selected after comparison with an 8B model, but quantified performance gains have not been disclosed.
Dexory has added NVIDIA Cosmos-Reason2-2B to its Storage Health software to investigate possible problems identified in robot-captured warehouse images. The 21 October integration extends the image analysis used in DexoryView, particularly when existing detection models return a low-confidence result. It is a software development for inspecting racks and goods, separate from the company’s Halos testing for autonomous robot safety.
Autonomous robots moving through warehouse aisles capture high-resolution images of goods stored at different rack heights. Storage Health examines those images for visible irregularities, including open cartons, pallets sitting out of alignment and items protruding from expected storage positions. Those conditions can be difficult to inspect from floor level in densely filled warehouses. The system identifies candidates for further review; an image alone cannot determine whether every flagged condition breaches a site rule or warrants a physical intervention.
Cosmos-Reason2-2B supplies an additional stage of visual interpretation when the existing models are uncertain. Dexory says natural-language prompts allow the foundation model to assess storage-specific conditions and contribute to the final classification. This is different from applying a single confidence threshold and reporting every uncertain detection as an exception. The output still depends on the quality of the image and the model’s interpretation; the company has not established that the reasoning stage removes the need for human checks.
Dexory evaluated the 2B and 8B versions of Cosmos-Reason2 before selecting the smaller model for production. Its stated criteria were accuracy and deployment efficiency. The model was fine-tuned using Dexory’s own warehouse imagery and tested against data from several warehouses. The company has not released a numerical comparison of precision, recall or false-alarm rates, nor enough detail about the test sample to determine whether one model is more accurate across every storage configuration.
The smaller model may require less computational resource than the 8B alternative, but Dexory has not specified how inference is distributed between robots and back-end systems. It has also not published processing latency or the number of images evaluated per scan. The operational change is the addition of a reasoning stage to Storage Health’s image-assessment process; the evidence does not establish that all analysis now runs on the autonomous vehicle without a network connection.
A warehouse rack presents challenges that are different from a clean product-image dataset. Goods may be partially hidden behind uprights, labels and stretch wrap can obstruct visible surfaces, and items are not always stored in precisely the same orientation. A box resting partly outside its expected footprint may be visible from one aisle position but difficult to judge from another. Multiple inspection passes and contextual reasoning can help distinguish normal variation from a condition needing review, although results still depend on image quality and model performance.
A flagged open box does not reveal why it was opened or whether the contents are complete. A pallet appearing out of alignment may need a physical inspection before anyone can judge its stability, particularly where the view is partly obstructed. Storage Health supplies observations for warehouse staff to assess, rather than moving goods or automatically changing stock records. No autonomous physical remediation or unsupervised inventory adjustment was included in the announced integration.
Repeated robot scans can build a more consistent record of conditions across high and densely occupied rack locations. Their value depends on whether the same positions are visible over successive passes and whether exceptions reach the relevant operating team in time for investigation. Poorly prioritised warnings can create additional workload, particularly if one condition appears in multiple images. Dexory has not disclosed post-integration alert volumes, staff review time or the proportion of warnings confirmed as actionable.
The development also raises the importance of maintaining consistent reference information. The interpretation of an image can change with the racking layout, pallet type, product packaging and normal storage rules for a particular location. Fine-tuning on warehouse imagery is one response to that variability, but performance at additional facilities must still be assessed. The announcement does not claim that a model trained on selected sites performs equally well in every possible warehouse environment.
Dexory is pursuing a separate NVIDIA Halos programme to explore safer perception and onboard autonomy for its mobile robots. Storage Health instead operates on the information the robots collect about surrounding stock. The overlap is the data-collection fleet and computing partner, not the technical purpose: a robot can navigate safely without detecting every misplaced pallet, while successful image classification does not certify its ability to avoid a moving forklift.
Dexory says the Cosmos integration broadens the storage conditions its software can recognise. The company has not quantified the resulting reduction, if any, in damaged stock, inspection time or safety incidents. Comparable data on detection accuracy, false positives and the time needed to investigate exceptions would be needed to assess the operational improvement across customer warehouses.
