CEL robots share successful practice to improve phone packaging

A fleet of 96 robots supplied practice data that helped two separate test robots pack phones successfully in 118 of 120 attempts. The Collective Experience Learning project from Tsinghua University, AgiBot and Infinigence AI asks whether machines that already have basic skills can improve by learning from one another, rather than requiring an expert to keep demonstrating corrections. The preprint first appeared on October 19, 2026; these are research results, not an unattended factory deployment.

A misplaced cable can disrupt a longer sequence. If every improvement requires another expert demonstration, more machines also mean more work for their teachers.

CEL works rather like a shared collection of worked examples. When an attempt fails, the system uses preceding camera observations and joint states to search successful robot experiences for comparable situations. It gives those useful action segments greater emphasis during training, then distributes an improved policy back to the fleet. People still label outcomes and reset the working environment.

Shared robot learning has a substantial history. Google described parallel grasp learning with 14 arms in 2016. The SOP system, introduced in January 2026, also connected fleet execution and human interventions to a common learner. CEL's particular approach is to use failure-guided retrieval of successful actions after expert-supervised initialization, without collecting additional expert demonstrations or corrective actions during that further learning stage.

In the phone-packaging experiment, success improved from 79.2% to 98.3%. With 120 trials in each evaluation, that means failures declined from 25 to two. Testing used two same-model robots excluded from training-data collection. Each task had its own trained policy, rather than one universal household worker.

The infrastructure matters, too. Robots need charging and sometimes disconnect. The team arranged experience collection so machines could submit their work independently and continue when a local fault interrupted another robot, rather than forcing the entire fleet to wait.

The practical question is whether more robots must mean proportionally more teachers. Broader tasks, different robot types and sustained deployment remain further tests.

Original preprint published: October 19, 2026, 11:41 Beijing time / October 18, 2026, 11:41 p.m. EDT.