Amazon's Robin arm learned to pick from 394,000 of its own deployed grabs

Robin is Amazon Robotics' first in-house robotic arm for sorting individual mailer envelopes and boxes out of cluttered piles: eight suction cups, a top-down camera view, and software that ranks possible picks. It began testing in 2021 and debuted at a Lakeland, Florida facility in 2022. The first strategies were hand-coded, and each failed in a characteristic way — one prioritised larger packages and stalled when smaller ones lay on top of or overlapped them; another favoured packages near the top of piles and kept missing buried, unreachable picks.

The fix was to let the fleet teach itself. Amazon's team trained a machine-learning model on more than 394,000 picking actions logged from deployed arms to estimate each candidate action's probability of success, then used an online pick ranker to choose what Robin should do. In the team's 2023 paper, the best ML method picked correctly 93.73% of the time against 92.28% for the best traditional algorithm — a 1.45-point gap that meant roughly 16,500 fewer failed picks over a work period of more than 1 million picks.

A follow-up experiment mimicking Robin fleet operations cut the failed-pick rate by 20% over 2 million picks, presented at the International Symposium on Experimental Robotics in July 2025. Robin's lessons carried into Cardinal, an arm that densely packs 50-pound boxes into carts ('think of it as a game of Tetris'), and the same foundation-model approach spread to Sparrow, which handles hundreds of millions of product types with no per-item hand coding.

Log every grab a deployed robot arm makes, train a model to estimate each candidate action's success odds, and let it rank actions — the fleet labels itself.

Every hand-crafted rule optimised one failure mode into another: prioritising big packages stalled on overlaps, prioritising top picks stranded buried ones.

The arms were already generating labelled data — every grab either worked or didn't — so training on 394,000 logged actions needed no extra labelling effort.

Per-pick differences look trivial until multiplied: Amazon judged strategies on failures per million picks, where 1.45 points equals 16,500 saved grabs.

At fleet scale, tiny per-action wins compound: a 1.45-point pick-success edge is 16,500 fewer failures per million picks — and your own operations log is the training set.

Robin became the template for Amazon's in-house arm line: Cardinal (dense cart packing, first run in Nashville in November 2022), Sparrow (foundation models to identify and grasp hundreds of millions of products, eliminating the need to code every item manually) and Vulcan, the first to add touch — force feedback sensors — to vision for picking inside fabric-covered pods. Amazon's CFO said on the July 2026 earnings call that the company expects to more than double its robotic-arm fleet this year; Amazon Robotics put current deployments in the single-digit thousands.