Amazon unveils Vulcan, its first robot with a sense of touch, to pick and stow about 75% of warehouse items

Amazon on May 7, 2025 unveiled Vulcan, a warehouse robot the company describes as its first with a sense of touch, at its Delivering the Future event in Dortmund, Germany. Vulcan uses force feedback sensors in its grippers to pick items from, and stow items into, the fabric storage pods that hold inventory in Amazon fulfillment centers. Amazon said the system can already handle about 75% of the types of items stored in those buildings, at speeds comparable to its front-line employees.

At launch Vulcan was working at two sites, the GEG1 fulfillment center in Spokane, Washington, and a facility in Hamburg, Germany, according to Amazon. A 2025 research paper published on Amazon Science reports that the stowing system had completed more than 500,000 stows in a large e-commerce fulfillment center, reaching human levels of packing density and speed. A companion paper says the picking system had operated in a live warehouse for more than six months and processed more than 12,000 customer orders. Amazon said it plans to deploy Vulcan systems over the next couple of years at sites throughout Europe and the United States.

Vulcan is aimed at the parts of the shelf that are hardest on people. Amazon's pods are divided into compartments about a foot square that hold up to 10 items on average, held in place by elastic bands. In Spokane and Hamburg the robot concentrates on the top rows, about 8 feet up, which normally require a worker to climb a step ladder, and on bins near floor level. Amazon Science said six Vulcan Stow robots had completed a pilot in Spokane, that 30 more were planned there for a beta phase, and that a larger deployment combining Vulcan Stow and Vulcan Pick was planned at the German site. People keep the middle shelves, densely packed bins and items such as fluid containers that need careful handling, the company said.

The hardware is built around touch. Every end-of-arm tool on Vulcan carries a six-axis force and torque sensor, according to Amazon Science, so the robot can gauge how hard it is pressing on an object and back off before the force becomes excessive. The stow tool, which Amazon likened to a ruler stuck onto a hair straightener, consists of two parallel plates with built-in conveyor belts and an extensible aluminium blade that nudges items aside to make room. The pick tool pairs a suction cup with a camera. A second arm with a hook moves the elastic bands, and three pairs of stereo cameras mounted on a tower build a three-dimensional model of each bin.

The software turns those signals into decisions. For stowing, three deep-learning models separate bands, compartments and items, trained in part on synthetic images in which generative AI added elastic bands; the system then searches for free space and chains control steps that Amazon Science names approach, extend blade, sweep and eject_item. For picking, a MaskDINO segmentation model sorts what it sees into four classes, separating non-items such as bands and bars from items that are accessible, lying below others or blocked by others, and a model trained by contrastive learning matches the target product against reference photos taken elsewhere in the building. During extraction the robot monitors suction pressure while its camera captures 10 low-resolution images per second. When Vulcan judges that it cannot move an item, it asks a human to take over, Amazon said.

Aaron Parness, Amazon's director of applied science, said the typical robot is "numb and dumb" and called Vulcan "a fundamental leap forward in robotics." He compared the approach to the way a person picks a coin off a table, touching the surface first and then sliding the fingers along it until they meet the coin. He told Amazon Science he saw "maybe 20 years of applications for this force-in-the-loop, high-contact style of manipulation."

Amazon said it had deployed more than 750,000 robots in its fulfillment centers over the previous dozen years and that robots play a role in completing 75% of customer orders. The stowing paper says that the diversity of items and strict business requirements for high rates and low defects had kept warehouse robots away from this task until now. Vulcan therefore targets work inside the bin itself, handling individual items in clutter.

Independent picking-robot developers have raised substantial venture money for related problems. Dexterity announced $140 million in equity and debt in October 2021, Nimble Robotics a $106 million Series C in October 2024, Covariant a $75 million Series C extension in April 2023, RightHand Robotics a $66 million Series C in February 2022 and Plus One Robotics a $50 million Series C in March 2023, according to their announcements. Those companies sell to many operators. Amazon builds Vulcan for its own network, which gives it a captive fleet of sites and a stream of training data from real orders.

For the wider logistics industry the point is the storage format. Fabric pods with elastic bands are cheap and dense, and they were designed around human hands. A robot that can work them by feel, instead of requiring rigid totes or rebuilt racking, lowers the cost of automating an existing building. Amazon has framed the benefit in safety terms, saying Vulcan lets employees work more in their power zone at about midriff height.

Several questions remain open. The 75% figure describes coverage of item types, not the share of a building's volume, and Amazon has not disclosed unit counts beyond the pilot and beta figures or defect rates in production. Each research result comes from a single site. Rollout speed will depend on whether hardware cost and reliability hold up across many buildings with different inventory.

On June 4, 2026, Amazon said that its investment of more than €10 billion over the coming years to expand and modernize fulfillment centers across Europe includes the expansion of Vulcan. The next markers are the number of sites running Vulcan by May 2027, the end of the two-year window Amazon set at launch, and whether the company publishes production defect and rate data.