Start with the work, not the robot
Warehouse work combines long-distance movement with lifting, picking, replenishment, packing, and loading. OSHA identifies repeated lifting and lowering, bending, overhead reaching, pushing, pulling, and awkward postures as ergonomic risks. It also describes warehouses as a moving mix of people, forklifts, and other equipment. That makes two useful starting questions: which movement is predictable enough to automate, and which manual exposure would the change actually remove?
A robot that moves pallets may reduce travel without solving piece picking. A picking arm may handle regular cartons while leaving damaged packaging, tangled items, unreadable labels, and substitutions to people. Define the job at the level of loads, routes, handoffs, and exceptions; the word ‘automation’ is too broad to be a requirement.
What is working now
Moving goods on structured indoor routes is established. NIST’s 2015 review describes automated guided vehicles used in factories and warehouses as tuggers, unit-load carriers, and automated fork trucks. The report is not a current market survey, but it documents a mature application pattern: repetitive material movement is a different problem from general-purpose mobile manipulation.
Scale is possible inside a tightly engineered operation. In June 2025, Amazon said it had deployed its one-millionth robot across more than 300 facilities. That is Amazon’s own operational report, not an independently audited performance benchmark, and it does not show that the same economics or reliability transfer to another warehouse. It does show that fleets doing bounded transport and handling jobs have moved beyond laboratory prototypes.
Where demonstrations and claims stop
Navigation is not a solved checkbox. A system needs localization, routing, traffic coordination, obstacle handling, docking, load transfer, and recovery when the map no longer matches the floor. NIST notes that poor routes, reduced speed around people, and unexpected pallets can limit efficiency. More recent NIST work still develops test methods for mapping, localization, docking, navigation, and object detection, which is evidence that capability must be measured in the intended environment.
Simulation can expose software failures cheaply, but it is not deployment evidence. NIST’s 2025 ARIAC challenge tests pick-and-place, assembly, and kitting in a simulated warehouse. Its results can compare algorithms under the challenge conditions; they do not establish safe behavior on a live floor, with real sensors, loads, people, dust, radio dropouts, or damaged inventory.
What building one involves
Begin with one route or handoff and record the real range of payloads, aisle widths, floor transitions, traffic, stopping distances, charging needs, and exception frequency. Specify interfaces to warehouse software and conveyors, then test throughput and recovery as well as ideal runs. Plan manual fallback and servicing from the start. OSHA says many robot accidents occur during programming, maintenance, testing, setup, or adjustment, when someone may enter the working envelope; a safe normal cycle is not a complete safety case.
Where Vesk fits
Simulation evidence for a bounded ROS 2 task.
Vesk can help generate and test ROS 2 Jazzy code in a controlled simulation before hardware work begins. That verification is simulation-only: it does not establish physical safety, sensor performance, payload behavior, hardware reliability, or readiness for a warehouse floor.