96 Robots Teaching Each Other: Tsinghua, Infinigence and AgiBot Push Collective Experience Learning to 97.5%

Three Chinese teams — Tsinghua University, AI infrastructure firm Infinigence AI and robot maker AgiBot — have jointly proposed a training paradigm called Collective Experience Learning (CEL), and validated it on a fleet of 96 real robots. The pitch is simple: instead of learning only from human demonstrations, robots should learn from each other's successes, so that every hard-won solution one machine finds becomes part of the whole fleet's repertoire.

The problem CEL targets is structural. Robots can be deployed by the hundreds, but experienced human operators cannot scale to match, and the current recipe — record demonstrations, correct failures by hand — makes human attention the bottleneck of embodied AI. If a robot could mine the experience of its peers the way a new hire studies a veteran's playbooks, deployment size would turn from a cost into an asset.

The validation covered five real tasks mixing home service and industrial work: flower arranging, folding clothes, charging a phone, assembling a stool and packing phones. According to the team's reported results, two robots that never took part in training ran 600 test trials and failed 15, a 97.5% success rate, and beat the average of six human operators on every one of the five tasks. Phone packing, a multi-step industrial task, rose from 79.2% to 98.3% success, holding at two failures across 120 consecutive runs. The team also reports a scaling trend: as the training fleet grew from 16 to 96 robots, success climbed from 86.7% to 98.3%. All of these figures are team-reported and have not been independently verified.

Underneath the results is a systems story that may matter as much as the algorithm. The fleet runs on FATS, a fault-tolerant asynchronous training system built on RLinf, an open-source large-scale reinforcement learning framework. Robots upload experience the moment a task ends rather than waiting for the fleet to sync, and the system is designed to absorb the realities of a hundred-machine deployment: in testing, the cluster collected 9,834 rounds of execution experience while 22 robots crashed 30 times between them, yet machine time lost to faults came to just 1.3%.

The algorithmic half is FERT, failure-guided experience retrieval fine-tuning. A single robot trial can run minutes and involve thousands of control steps, but the reward is often a single bit: success or failure. FERT handles that sparsity by learning from both sides of the ledger — copying what worked in similar past situations, and mining failures for what went wrong — so a robot that stalls retrieves the closest successful episode in the shared pool and imitates the difference.

The three-way split of labor is the other tell. Tsinghua supplied the research, Infinigence the training infrastructure, AgiBot the hardware and the real scenes. That is the pipeline embodied AI has been missing: methods that survive contact with an actual fleet rather than a lab bench.

If the reported numbers hold up under outside scrutiny, the implication is uncomfortable for the demonstration-data business: a fleet that teaches itself gets better the longer it runs and the bigger it gets, which is exactly the curve human demonstration datasets cannot follow. The next test is replication by a team with no stake in the result.