AgiBot unveils CEL with Tsinghua University and Infinigence AI, tests collective learning with 96 robots

Shanghai Zhiyuan Innovation Technology Co., Ltd. announced collective experience learning (CEL) with Tsinghua University and Infinigence AI on October 30, 2026. A fleet of robots, including 96 physical units, was evaluated on five tasks: flower arranging, clothes folding, phone charging, stool assembly and phone packaging. CEL's overall success rate was 97.5%. The official article was published on November 2, 2026.

CEL was presented as a system that shares successful operations across a fleet and uses failure scenarios to select what to learn next. AgiBot reported an overall success rate of 97.5% across five tasks and a 1.3% loss in operating time due to failures among 96 robots. The evaluation therefore concerns not only learning outcomes but also the ability to keep a large number of physical robots running.

What stands out about CEL is that the number of robots was not simply increased to expand the scale of the experiment; their task trajectories were treated as a learning resource to be shared and used to update the next policy. FERT narrows down learning targets based on observations of failed tasks and retrieves successful examples from a shared experience pool. FATS is designed to prevent an individual device going offline from disrupting the entire fleet. The reported gains in success rate should therefore be considered alongside the system for continuously operating physical robots, communications and computing resources, rather than as the result of the learning algorithm alone. An earlier report on AgiBot's announcement, AgiBot updates simulation platform “Genie Sim” to support simulation of flexible objects, covered simulation, evaluation and post-training. In the official account, CEL reuses successful and failed trajectories from real-world tasks across a fleet. The companies have not said that the two systems form a connected operational process, but simulation-based evaluation and experience sharing among physical robots can be distinguished as different stages of development. The formula described in AgiBot proposes “Worken,” a robot utilization metric, reports leaderobot.com combines the number of operating robots, autonomous work hours and an output-efficiency ratio. CEL's announcement reports task success rates under conditions involving a larger number of robots, but does not present the metric as a way to measure operating volume or output efficiency. The different measures should not be conflated: success rates and actual operating volume need to be evaluated separately. The evaluation covered five tasks, 600 physical-robot tests in total and about 30 robot-hours. The reported results do not establish that the same performance will apply to other tasks or long-term operation. In addition, the account of scaling up the number of robots includes both a comparison from 16 to 96 across the five tasks and one from 6 to 96 for phone packaging. The conditions and scope of the success-rate calculations for each comparison need to be clarified. Other questions to address include whether repeated trials produce similar results and how the quality and safety of shared trajectories will be managed.

CEL identifies situations to target for learning from observations immediately before a failure, then searches successful trajectories shared across the fleet for operations in similar situations. The updated policy is distributed to the robots, and new successful and failed trajectories are fed into the next round of learning. The article says CEL's success rate across five tasks exceeded the rates for SFT, at 36.3%, and HG DAgger, at 81.0%, as well as the average for six human teleoperators. The FATS learning infrastructure dynamically manages 100 robots, 50 edge nodes and six cloud servers. In an experiment with 96 robots, it collected 9,834 task runs. Thirty robot failures resulted in a loss of 1.3% of operating time across the fleet. The article says that for phone packaging, the success rate rose from 86.7% to 98.3% when the number of participating robots increased from six to 96. Elsewhere, it describes an increase from 16 to 96 robots and a success rate rise from 86.7% to 98.3%, but it is not clear whether the two statements cover the same scope.