Warehouse picking robots: can they grab any item?

Warehouse picking robots are machines that reach into a bin or shelf, pick out one product and place it in an order container. Online shops need this because each order has to be assembled item by item. Products differ in shape, weight, packaging and softness, so the robot has to work out where to grip and how hard to press. The published evidence reviewed for this article is limited, and most of the larger figures come from the companies themselves.

What has been demonstrated

The clearest published result is a 2019 paper in Science Robotics by Jeffrey Mahler and colleagues at UC Berkeley. The paper explores grasping with two or more different grippers. Its Dex-Net 4.0 policies, for a parallel-jaw gripper and a suction cup, were trained on 5 million synthetic depth images, grasps and rewards. The paper reports that, on a physical robot with two grippers, the policy consistently cleared bins of up to 25 novel objects with reliability above 95% and at more than 300 mean picks per hour (paper abstract). This was bin clearing in a test setting. It does not show error rates across a live product range, and the abstract does not state any limitations.

A narrower result comes from a competition. A paper on the Cartman robot, first posted in 2017 and listed as due to appear at the ICRA 2018 conference, says it won first place in the finals of the Amazon Robotics Challenge by stowing 14 of 16 items and picking all 9 items in 27 minutes, for 272 points (arXiv). This was one timed competition run.

What companies and their researchers report

The figures in this section are claims made by companies or by the teams that built the systems. None has been independently tested in the sources reviewed.

• Amazon Sparrow. Amazon describes Sparrow as able to recognise and handle "millions of items" using computer vision and AI (Amazon). The announcement gives no accuracy figures, unit counts, facility names or rollout timeline.
• Amazon Vulcan (2025). Amazon says Vulcan can pick and stow approximately 75% of the item types it stores at its fulfilment centres. It uses force-feedback sensors and, for picking, an arm with a camera and a suction cup. When it cannot move an item it can ask a human partner to take over. Amazon says it works at speeds comparable to front-line employees, but gives no figures. The page does not explain how the 75% was measured and gives no error rate. At Amazon's fulfilment centres in Spokane and Hamburg, Vulcan is focused on the top rows of storage pods, which typically need a step ladder, and Amazon says it plans to deploy Vulcan systems "over the next couple of years, at sites throughout Europe and the United States" (Amazon).
• Stowing preprint (2025). A 2025 arXiv preprint on a robot that stows items into fabric pods reports over 500,000 stows in a large e-commerce fulfilment centre. It says the system "achieves human levels of packing density and speed" and prioritises overhead shelves to improve the safety of people working alongside it (arXiv). These are the authors' own figures in a preprint, not an independent test, and the abstract gives no success or defect rate. Stowing, which puts items into storage, is related to picking but is a different task.
• Ocado (2020). Grocery Dive reported in November 2020 that Ocado was spending $287 million to buy Kindred Systems ($262 million) and Haddington Dynamics ($25 million) to speed up its robotic picking technology. According to Ocado's announcement, as reported there, the robotic arms at the five picking stations at its Erith site in the UK had doubled their pick-and-pack rate. The article gives no baseline, no new figure and no accuracy figure (Grocery Dive).
• Covariant (2024). Covariant announced RFM-1, which its CEO described as a large language model for robot language. According to TechCrunch, it draws on data collected from the deployment of Covariant's Brain AI platform. The report describes a live demonstration of commands such as "pick up the red object", but gives no pick-success, accuracy or throughput figures. The CEO said the company expects RFM-1 to work with a majority of the hardware on which Covariant software is already deployed. That is an expectation, not a measured result (TechCrunch).

Limits and open problems

• Varied items. Suction cannot seal on porous items such as clothing, and parallel-jaw grippers may not reach both sides of some tools and toys (UC Berkeley News). Mahler is quoted there as saying "Any single gripper cannot handle all objects". Amazon does not say what the remaining 25% of item types involve.
• Damaged goods. None of the sources reviewed for this article report how picking systems handle damaged or defective products.
• Speed against people. The human comparisons are Amazon's statement about Vulcan's front-line speed, which has no numbers, and the stowing preprint's claim of human-level speed, which the abstract does not quantify. The 300 picks per hour figure comes from a bin-clearing test and cannot be set against a warehouse worker's rate.
• Cost. The Ocado deal price is the only cost figure reviewed. It is an acquisition price, not the cost of installing or running robots in a warehouse.
• Safety around staff. The sources reviewed do not describe safety arrangements for robots working near people in detail. The related details are Vulcan's handover to a human partner, its focus on pod rows that would otherwise need a ladder, and the stowing preprint's focus on overhead shelves.

Three states at a glance

Demonstrated, under test conditions: the Dex-Net bin-clearing results and the Cartman competition run.

Reported, not independently verified (company or press claims, or self-reported preprint figures): Amazon's Sparrow and Vulcan claims, including the 75% coverage and speed claims, the 2025 preprint's stow count and human-level claims, Ocado's doubled pick-and-pack rate and Covariant's RFM-1 claims.

Speculative: that if machines could pick reliably, warehouses might move staff away from repetitive reaching and bending; that soft goods could need grippers that combine different gripping methods; and that shared grasp data across a network of warehouses could let smaller sites use picking robots without custom training for each product range. None of these has been shown in the sources reviewed.

What to watch next

• Whether Amazon or Ocado publish pick success or error rates for named sites and item types.
• Whether Vulcan is deployed beyond Spokane and Hamburg, which Amazon's plan of the "next couple of years" (stated in 2025) would imply.
• Whether the remaining item types, such as soft or porous goods, are picked without a human handover.
• Whether independent benchmarks on unseen products report success rates in warehouse conditions.
• Whether Ocado publishes absolute pick rates or accuracy figures for its Erith stations.