A look at Covariant.AI, which builds warehouse robots trained with reinforcement learning, as few warehouse tasks remain that cannot be automated
and say they're ready for the big time. https://www.technologyreview.com/ ... @techreview : “Every time, we expected that they would fail with the next product, because it became more and more tricky ... But the point was they succeeded, and everything really worked. We've never seen this quality of AI before.” https://www.technologyreview.com/ ... https://twitter.com/... Karen Weise / @kyweise : .@CadeMetz and @satariano on the big AI and robotics breakthrough in warehouse work: picking through a bin of random items https://www.nytimes.com/... Isaac Saul / @ike_saul : A robot arm is successfully sorting boxes in a factory with 99% accuracy. Might not seem like much, but it's a task that has been out of reach of the most advanced artificial intelligence. Speaks to the threat of automation @AndrewYang often discusses. https://www.nytimes.com/...
Context & Ripple Effects
This 2020 profile captures Covariant at the moment bin-picking — grabbing arbitrary items from jumbled totes — stopped being the wall of warehouse automation, with the company's reinforcement-learning arms sorting boxes at roughly 99% accuracy and winning over observers who expected each harder product to break them. The arc since then validates the bet: Covariant later shipped RFM-1, a robotics foundation model that reasons about physics and reduces the need for bespoke programming per task.
The competitive frame matters too: Amazon was building the same capability in-house, and by 2025 its Sparrow, Cardinal, and Proteus robots were taking over selecting, picking, and cart-carrying roles — while the pandemic pushed techniques like these beyond Amazon to smaller American retailers scrambling to automate fulfillment.
First-order effects
- Warehouse operators gain a vendor for the hardest remaining manual task — mixed-item picking — without writing custom code per SKU or facility, since Covariant's models learn from real deployments.
- Amazon's in-house robotics program now has a direct external rival selling the same picking capability to every other logistics operator.
Second-order effects
- Shippers like FedEx and mid-sized retailers become buyers of vision-equipped robot arms rather than only Amazon-scale players, widening the market beyond one giant's captive fleet.
- Robotics vendors shift from selling task-specific machines to selling generalist models trained on accumulated deployment data, making dataset scale the moat.
Third-order effects
- If the pattern holds, warehouse work consolidates around a handful of AI-model owners whose fleets improve with every pick, squeezing integrators who program robots task-by-task.
- Labor planning in fulfillment shifts structurally: roles like picking move from human-staffed lines to supervised automation, with the pace set by model generality rather than per-site engineering.
The trend: Warehouse robotics is moving from bespoke, single-task machines to generalist AI models trained on live deployment data, with Covariant's picking breakthrough as an early proof point and Amazon's fleet as the scale endgame.