Remy AI Is Building Warehouse Robots That Can Actually Pick Things Up

The Macro: Warehouses Are Still Mostly Manual

The warehouse automation market is projected to exceed $40 billion by 2028. But walk into most warehouses today and you will find humans doing the vast majority of the work. Picking items off shelves, packing boxes, sorting products, and handling the thousands of different SKUs that modern e-commerce demands.

The reason is dexterity. Moving boxes on conveyors is solved. Autonomous mobile robots from companies like Locus Robotics and 6 River Systems can transport bins around a warehouse. But actually picking up a specific item from a cluttered shelf, orienting it correctly, and placing it in a box requires a level of manipulation skill that most robots cannot match.

The items in a warehouse vary enormously in size, shape, weight, texture, and fragility. A robot that can pick up a hardcover book cannot necessarily pick up a bag of chips, a bottle of shampoo, or a bundle of cables. Each item requires different grasp strategies, and the robot needs to figure out the right approach in real time.

This dexterity problem is why companies like Amazon still rely on hundreds of thousands of human workers in their fulfillment centers despite investing billions in robotics.

Remy AI, backed by Y Combinator, is attacking this problem head-on with robots designed for high-dexterity warehouse tasks and a proprietary model that learns new manipulation skills with less training data than standard approaches.

The Micro: Oxford PhDs Bringing Dexterity to the Warehouse

Oscar Brisset (CEO) is an Oxford graduate and former BCG consultant who advised Fortune 500 logistics companies. Ben Kaye (CTO) holds an ML PhD from Oxford specializing in computer vision and robotics, and previously engineered safety-critical systems at OrganOx. The combination of logistics industry knowledge and robotics research expertise is a good fit for this problem.

The technical differentiation is in the training pipeline. Vision-language-action (VLA) models are the current standard approach for teaching robots new tasks, but they require large amounts of demonstration data. Remy AI claims their custom training pipeline learns new tasks with less data than vanilla VLA fine-tunes. This matters because collecting robot training data is expensive and time-consuming.

The hardware design is optimized for demanding warehouse environments. Warehouses are dusty, temperature-variable, and run 24/7. Consumer-grade robotic components fail quickly under these conditions. Remy’s hardware is built to operate robustly in these environments.

Competitors include Covariant (AI-powered robotic picking), RightHand Robotics (piece-picking), and Dexterity (palletizing and depalletizing). The warehouse robotics space is active but still far from solving the general picking problem. Remy’s claim of learning new tasks efficiently could be a meaningful advantage in a world where warehouses handle thousands of different products.

The Verdict

Remy AI is working on one of the hardest and most valuable problems in robotics. Warehouse dexterity is the bottleneck that prevents full automation of fulfillment operations.

At 30 days: how many different product types can Remy’s robots reliably pick, and what is the pick success rate compared to human workers?

At 60 days: how quickly can the system learn to handle a new product category? Hours of training data versus days versus weeks determines practical utility.

At 90 days: are warehouse operators seeing measurable throughput improvements and cost savings from Remy deployments?

I think Remy AI is tackling the right problem at the right time. The labor shortage in warehousing is real, and the dexterity gap is the specific bottleneck. If their training pipeline genuinely requires less data than competitors, they can adapt to new customers and products faster, which is the key to scaling a warehouse robotics business.