Warehouse picking robots
Also known as Robotic picking, Piece-picking robots, Item picking robots
Warehouse picking robots grab individual products from bins or shelves so that orders can be packed item by item. Amazon, Ocado and AI companies are deploying them, but the public performance figures so far come mostly from company announcements and controlled lab tests.
In plain language
Warehouses hold millions of different products, and online orders often need just one or two items at a time. A picking robot reaches into a bin or shelf, grabs a single item and places it in an order container. This is harder than it sounds, because products differ in shape, weight, packaging and how soft or slippery they are, so the robot has to work out where to grip and how hard to press.
Companies take different approaches. Amazon's Sparrow uses cameras and AI to handle individual products, and its Vulcan robot adds force-feedback sensors so it can feel how hard it is pushing or holding an item. Ocado, the UK online grocer, bought the robotics firms Kindred Systems and Haddington Dynamics in 2020 to speed up the commercial delivery of its robotic picking systems. Most public performance figures come from company announcements, so they should be read as claims rather than independent test results.
Technical explanation
A picking system combines perception (cameras or depth sensors that estimate where an item is), grasp planning (choosing a suction point or a pinch point) and control (moving the gripper and sensing contact). Suction grippers are widely used because they handle many items, but a suction cup cannot seal on porous objects such as clothing. Mahler and colleagues argued that no single gripper type can handle every object. Their Dex-Net 4.0 policy combined a parallel-jaw gripper and a suction gripper and was trained in simulation on 5 million synthetic depth images, grasps and rewards.
In the paper's reported experiments, this policy cleared bins of up to 25 novel objects with reliability greater than 95%, at more than 300 mean picks per hour. These are controlled bin-clearing results on a specific hardware setup. They do not show error rates across a live warehouse catalogue, or how often a pick needs human help. Vendor descriptions of deployed systems, such as Amazon's Vulcan, instead report item-coverage shares and do not state error rates.
Newer work tries to generalise across items with large learned models. Covariant announced its RFM-1 model in March 2024, describing it as the result of, among other things, data collected from deployments of its Brain AI platform. Press coverage reported no pick-success or accuracy figures for it. The 2020 Grocery Dive report says Ocado has doubled the pick-and-pack rate of the robotic arms at its Erith picking stations, but it gives no absolute rates or accuracy figures.
What the evidence says
Each statement is typed by the strength of its evidence and linked to the source that supports it.
Demonstrated Shown in a published experiment, test or shipping product.
Amazon's Sparrow announcement gives no accuracy figures, unit counts, facility names or rollout timeline.
Sources: [1] · Full announcementIn the paper's reported experiments, a combined parallel-jaw and suction policy trained on 5 million synthetic depth images cleared bins of up to 25 novel objects with reliability greater than 95% at more than 300 mean picks per hour.
Sources: [3] · Abstract · medium confidenceA single gripper type cannot handle every object, and a suction cup cannot create a seal on porous objects such as clothing.
Sources: [4], [3] · Quoted explanation of gripper choiceOcado agreed to acquire Kindred Systems for $262 million and Haddington Dynamics for $25 million, a total of $287 million, announced on 2 November 2020 to speed up the commercial delivery of its robotic picking systems.
Sources: [6] · Deal summaryReported, unverified Stated by a company or team but not independently confirmed.
Amazon says Sparrow can recognise and handle millions of items, using computer vision and AI.
Sources: [1] · Announcement text · medium confidenceAmazon says Vulcan can pick and stow approximately 75% of the various types of items it stores at fulfilment centres, working at speeds it describes as comparable to front-line employees.
Sources: [2] · Announcement text · medium confidenceAmazon says Vulcan uses force-feedback sensors that tell it how hard it is pushing or how firmly it is holding an item, and a camera that checks it has grabbed the right item, and that when it cannot move a specific item it can ask a human partner to take over.
Sources: [2] · Sensing and handoff sections · medium confidenceCovariant says RFM-1 is the result of, among other things, data collected from deployments of its Brain AI platform, and that it has deployed robots at warehouses with success. Press coverage reported no pick-success or accuracy figures for the model.
Sources: [5] · Training data and deployment statementsThe article reports that Ocado has doubled the rate at which robotic arms at five picking stations at its Erith, UK facility can pick and pack. No absolute pick rate or picking accuracy figure is given.
Sources: [6] · Operations statement · medium confidenceProposed A design or plan that has not yet been built or tested.
Amazon names Spokane (GEG1) and Hamburg as sites focused on Vulcan's use in the top rows of storage pods, and plans to deploy it at sites throughout Europe and the United States over the next couple of years.
Sources: [2] · Deployment section · medium confidenceLimitations and open problems
- Amazon's 2022 Sparrow announcement gives no accuracy figures, error rates, deployed unit counts or facility names.
- Amazon says Vulcan can pick and stow about 75% of the item types stored at its fulfilment centres. The announcement does not say what the remaining item types involve or how often a pick fails.
- Amazon says that when Vulcan cannot move a specific item, it can ask a human partner to take over, so the sources do not claim full automation of difficult picks.
- The Dex-Net 4.0 results come from policies trained in simulation and tested on bin-clearing tasks with up to 25 novel objects. The abstract does not state limitations.
- Suction cannot seal on porous items such as clothing, and parallel-jaw grippers may not reach both sides of some tools and toys (Mahler, quoted by Berkeley News, 2019).
- Ocado's picking accuracy was not found. The only rate figure is a relative one, that its Erith picking stations have doubled their pick-and-pack rate, and the 2020 acquisition announcement gives no absolute performance figures.
Possible future uses
Ideas, not predictions. None of these has been demonstrated.
- Reliable handling of clothing and other soft goods could depend on grippers that combine suction with fingers, as the Dex-Net work suggests.
- Robots that learn from their own failures across many sites could steadily widen the share of item types they can pick.
- Shared grasp-learning data across fulfilment networks might let smaller warehouses use picking robots without custom training for each product range.
Timeline
- 16 Jan 2019Mahler and colleagues publish the Dex-Net 4.0 ambidextrous grasping policy in Science Robotics, combining suction and parallel-jaw grippers. Sources: [3], [4]
- 2 Nov 2020Ocado announces it will acquire Kindred Systems ($262 million) and Haddington Dynamics ($25 million) to speed up the commercial delivery of its robotic picking systems. Sources: [6]
- Nov 2022Amazon introduces Sparrow, a system that identifies and handles individual products using computer vision and AI. Sources: [1]
- 11 Mar 2024TechCrunch reports that Covariant has announced RFM-1, a robot foundation model. Sources: [5]
- May 2025Amazon unveils Vulcan, a robot that uses touch sensing to pick and stow items, and names Spokane and Hamburg as sites for deployment. Sources: [2]
Futures this could help bring about
Related breakthroughs
Glossary
- Pick rate
- The number of items or orders a robot or worker picks per hour. Published figures often do not say whether they include retries or human help.
- Suction gripper
- A gripper that holds an item by pulling air out through a cup. It works well on smooth, sealed surfaces but poorly on porous items.
- Parallel-jaw gripper
- A gripper with two opposing fingers that pinch an object. It can struggle to reach both sides of some items.
- Item coverage
- The share of a warehouse's product types that a robot can pick. Coverage figures do not tell you how reliably the robot picks each type.
Sources
- company announcement
Amazon introduces Sparrow, a state-of-the-art robot that handles millions of diverse products
Amazon (About Amazon), Nov 2022. Accessed 21 Oct 2026. - company announcement
Amazon Vulcan robot pick and stow touch announcement
Alex Davies. Amazon (About Amazon), May 2025. Accessed 21 Oct 2026. - peer reviewed paper
Learning Ambidextrous Robot Grasping Policies
Jeffrey Mahler, Matthew Matl, Vishal Satish et al.. Science Robotics, 16 Jan 2019. Accessed 21 Oct 2026. - university press release
'Ambidextrous' robots can learn to pick up anything
UC Berkeley News, 16 Jan 2019. Accessed 21 Oct 2026. - news
Covariant is building ChatGPT for robots
Brian Heater. TechCrunch, 11 Mar 2024. Accessed 21 Oct 2026. - news
Ocado acquires pair of firms to speed up robotic picking tech
Sam Silverstein. Grocery Dive, 3 Nov 2020. Accessed 21 Oct 2026.
Revision history
3 changes since this profile was created.