Google trained a fleet of 23 robots to sort office trash

Sorting office trash properly is something even humans struggle with every day. Google decided to let reinforcement learning deal with the mess instead.

Researchers from the Google Brain team revealed results from a two-year experiment where they deployed a fleet of 23 mobile manipulator robots into real Google office buildings to sort waste and recyclables.

The roving janitor fleet

Instead of sitting on an assembly line, these robots roamed around hallways looking for waste stations consisting of compost, recycling, and landfill bins.

Once a robot approached a station, its job was to look inside, identify misplaced items, and move them where they belonged. Bottles and cans went to recyclables, paper cups and cardboard food containers went to compost, and everything else stayed in landfill trash.

Sorting crumpled cups and wet containers sounds mundane, but in robotics, picking up arbitrary squished objects in shifting lighting is notoriously difficult.

To pull this off, the team built on their QT-Opt reinforcement learning framework and trained the bots across four distinct layers. They started with basic scripted policies to gather early data, then scaled up in a physics simulator using RetinaGAN to bridge the sim-to-real visual gap.

540,000 runs in a robot classroom

Learning entirely on the job in live offices is too slow because trash volumes fluctuate heavily by the day. To solve this bottleneck, Google set up a dedicated robot classroom.

Inside the classroom, 20 robots practiced continuously on representative waste stations, churning out 540,000 trials. Meanwhile, other robots collected 32,500 trials from live deployment across 3 office buildings and 30 waste stations.

Across 4,800 evaluation trials spanning 240 bin configurations, performance climbed steadily as more data rolled in. The final system sorted objects with an average accuracy of 84%.

More importantly, real-world deployment data logged between 2021 and 2022 showed that the robots reduced waste bin contamination by 40% to 50% by weight.

Why it matters

Deep reinforcement learning has spent years shining in simulated environments, from chess and Go to video games, but it rarely survives contact with the messy physical world.

Google showed here that real-world mobile manipulation can actually generalize across diverse, unscripted environments if you feed it the right hybrid curriculum. The combination of synthetic data, dedicated robot classrooms, and on-site fine-tuning provides a practical blueprint for training service robots at scale.

Nobody is going to buy a fleet of these robots tomorrow just to fix office recycling. The hardware cost and operational overhead are still far too high for everyday commercial viability.

Still, we think Google's hybrid approach of pairing a high-throughput robot classroom with live field testing was a brilliant way to dodge real-world data bottlenecks. If reinforcement learning can reliably cut contamination rates in noisy corporate cafeterias, it brings autonomous material handling and warehouse sorting one giant step closer to the real world.