IN Brief:
- Dexory is testing NVIDIA Halos with an IGX Thor-based architecture for autonomous mobile robots.
- The intended single-board design would combine safety functions and wider onboard perception.
- Dexory has not disclosed completed fleet deployment, a certification outcome or measured safety improvement.
Dexory is testing the NVIDIA Halos safety platform as it develops a new computing and perception architecture for autonomous warehouse robots. The company is exploring how NVIDIA IGX Thor and the Halos safety stack could support a single-board design for machines operating among staff, forklift trucks and other equipment. Dexory announced the work on 21 October and intends to extend the technology across its international fleet, but has not said that deployment or safety certification is complete.
The proposed architecture brings AI processing and functional safety capabilities into the robot’s onboard computing arrangement. Dexory calls its objective ‘inside-out safety’, referring to the machine’s ability to perceive the operating environment around itself rather than relying solely on fixed infrastructure. The company has not specified the final sensors, the safety-related control architecture or the extent to which existing machines would require hardware replacement. The single-board design remains a development objective, not a confirmed specification for all deployed robots.
As a robot moves along a warehouse aisle, it must account for stationary racking as well as obstacles that appear or move unexpectedly. People may cross its route, manually driven trucks may reverse into an aisle and pallets may be left where an earlier map showed open space. A navigation plan alone does not resolve those changing conditions. Sensors, local processing, motion controls and site traffic rules each contribute to how a machine detects an obstruction and reacts before proceeding.
Functional safety is a separate discipline from general AI perception. A vision model may classify an object correctly in ordinary conditions but still produce uncertain or incomplete results when lighting changes, part of the scene is obscured or a sensor becomes unavailable. A safety-related system must also account for faults and specify how the machine reaches or maintains an appropriate safe state. Dexory has not reported a named conformity assessment, a completed safety case or independently verified reductions in collisions arising from the Halos tests.
NVIDIA positions Halos as a safety platform combining real-time AI, functional safety and edge computing with IGX Thor. Processing on the robot can reduce dependence on a remote connection when decisions must be made during movement. That design principle does not establish an achieved reaction time or a particular degree of fault tolerance. Those properties depend on the full sensor, controller and actuation chain, and Dexory has not published measured latency, redundancy or fault-handling results for the configuration under test.
Dexory’s machines work in logistics and manufacturing sites, collecting data used by its DexoryView platform to give operators visibility of inventory and storage conditions. Their established operating purpose is not identical to that of a forklift or a robotic manipulator. The safety design must reflect the movements, speeds, sensor coverage and foreseeable interactions of the particular device, rather than assuming that requirements for all autonomous warehouse equipment are interchangeable.
Dexory says the new approach should expand the robot’s perception of activity close to it. Wider perception could improve the information supplied to navigation and obstacle-handling functions, although the operational effect will be determined by sensor coverage, how uncertain detections are handled and which control actions the robot is permitted to take. No before-and-after detection benchmark or final safety-rated control boundary has been released. An assertion that the system has eliminated blind spots would be unsupported.
Several warehouse operating conditions place pressure on perception systems. Pallet racks and equipment can create occlusions, reflective surfaces may affect certain sensors, and mixed traffic makes encounters hard to predict from static maps. Lighting and changes in the position of stock can also affect some visual approaches. Testing a proposed safety architecture needs to consider these conditions, but Dexory has not released a test matrix or independent results establishing performance across them.
A fleet-wide update would extend beyond installing a new processor board. Site configurations and robot software would need controlled validation, with maintenance procedures and operator training aligned to the revised architecture. Local traffic arrangements, pedestrian crossings and emergency response procedures also influence residual risk. The company has not disclosed a deployment timetable, a count of robots in the tests or a programme for upgrading units already operating at customer sites.
The Halos programme concerns the safe movement of Dexory robots, whereas the company’s separately announced NVIDIA Cosmos integration concerns images those robots capture of stored goods. In Storage Health, the model supports classification of possible stock and rack anomalies. Halos is being evaluated for onboard autonomy and perception. The same fleet can contribute data to an inspection service without the inspection model establishing that the moving machine satisfies functional safety requirements.
Dexory’s confirmed milestone is a testing programme using NVIDIA Halos and an intention to develop the architecture for wider use. Evidence of fleet deployment will require a disclosed configuration, completed validation and evidence from operating conditions representative of warehouses with mixed traffic. Until then, improvements in detection coverage and safe autonomy remain design objectives rather than measured outcomes.
