How Amazon taught a warehouse robot to move things by touch

Amazon’s Vulcan robots address a familiar physical problem: reaching into a crowded shelf without pulling out everything beside the object you want. The storage compartments have flexible fabric walls and elastic bands across the front. Moving one item can change the position of several others.[1]

The system pairs cameras with force and torque sensors. Contact becomes information about the scene, helping the robot control how hard it pushes and when to change its movement. Amazon publicly described the system in May 2025, after warehouse pilots.[1]

Making room for one more item

Vulcan Stow uses a specialised gripper with small conveyor belts to slide a product into place. A separate tool moves the retaining bands, while an extendable blade can push existing contents aside. The important decision is whether there is usable room, including room the robot can create by rearranging what is already there.[2]

A learned selection model can estimate whether a proposed action will succeed and how long it will take. This allows the robot to choose between a straightforward insertion and a more involved movement. Touch feedback during a sweep also helps it revise its estimate of the available space.[2]

Choosing and extracting the right object

Vulcan Pick uses reference product photographs to identify the requested item. It then looks for a suitable surface for its suction tool and plans a route out of the compartment. Continuous visual feedback helps it respond if the object shifts during extraction.[3]

That selection process is part of the reported result. In the January–March sample, 1,270 of 6,561 requests were rejected because band-opening or picking could not be planned: 19.4%. A further 41 system failures and 93 band-operation failures left 5,157 extraction attempts. Of these, 4,690 succeeded. That is 90.9% of attempted extractions, or 71.5% of all requests assigned to this station. These denominators describe different stages; neither measures the share of all warehouse work the robot could handle.[3]

This gives the deployment a useful practical shape. A robotic station does not have to handle every possible shelf arrangement to take on a worthwhile share of the work. It does need a dependable way to recognise which work belongs elsewhere and return it to the existing process.

Why the shelves matter

Amazon’s technical account explains that items are distributed across pods to reduce competition between picking stations for the same storage unit. The robot is therefore adapting to a working logistics system, with a changing assortment of products, rather than receiving a carefully arranged tray of identical parts.[1]

In its June 2026 European robotics update, Amazon said Vulcan had expanded from Spokane, Washington, to more complex picking work at its Hamburg facility. Further European expansion was part of the announced programme. That is a company-reported deployment account, separate from the earlier papers’ measured trials.[4]

The same update situates Vulcan alongside other machines with different jobs: mobile Proteus robots transport carts, while STARK handles full totes. The division is instructive. Moving a load across a building and extracting one product from a flexible shelf require different combinations of sensing, hardware and control.[4]

Vulcan’s most interesting contribution is that it uses the consequences of an action to guide the next part of that action. Feeling objects shift makes a crowded compartment more understandable. That is a useful step towards robots working with the untidy arrangements people manage routinely.

Report an error or suggest a correction