Tsinghua, together with Infinigence AI and AgiBot, proposes Collective Experience Learning (CEL); validation on 96 humanoid robots achieves a 97.5% success rate

Related topics top start Related topics top end Futures promotion begin Futures promotion end Miaopai begin Miaopai end Video player start Video player end Market chart begin Market chart end

Recently, Tsinghua University, together with Infinigence AI and AgiBot, introduced a new paradigm called Collective Experience Learning (CEL), and completed validation on a group of 96 real humanoid robots. Simply put, this new method changes the previous model in which robots were mainly taught by humans hand by hand, allowing robots to learn from each other's experience: the pitfalls one machine has stepped into and the techniques it has figured out can be shared with all its companions. The more machines there are, the richer the accumulated experience and the stronger the overall capability, opening up a new approach for the large-scale deployment of humanoid robots. The Collective Experience Learning paradigm CEL: enabling robots to evolve from relying on human demonstrations to learning within a collective

In the past, training robots mostly relied on humans demonstrating actions and manually correcting errors. But if robots are deployed in large batches, the number of professionals who know how to operate them is limited, and manpower cannot keep up with the pace of robot expansion, becoming a real problem constraining the industry's development. CEL collective learning was born to solve this pain point: after robots complete tasks, they pool and share their successful experiences and failed lessons in a unified way. When a certain robot encounters a difficult problem, it can retrieve the past practical experience of its companions for reference and learning, without all of them having to wait for human guidance, achieving the collective evolution of the robot group.

In real-environment tests, the team set up five household and industrial tasks: flower arranging, clothes folding, phone charging, stool assembly, and phone packaging. After collective learning training, the robot group's overall task success rate reached 97.5%, and performance on all five tasks exceeded the average level of 6 human operators. Among them, the phone packaging task, which has a complex workflow, saw its success rate rise from 79.2% to 98.3%. The experiment also confirmed a scaling effect: as the robots participating in training expanded from 16 to 96, overall capability further improved accordingly.

This achievement is supported by both algorithms and systems. At the system level, a fault-tolerant asynchronous training system, FATS, was built; even with network fluctuations and hardware failures of some robots, nearly a hundred machines can still continue collaborative training, with time loss caused by hardware failures accounting for only 1.3% of total operating time. At the algorithm level, the FERT experience retrieval fine-tuning algorithm is adopted, learning operating methods from successful cases and also locating problems from failure records, helping robots quickly optimize action strategies. Infinigence AI is rooted in Xuhui and grew up in the Model Space

Among the companies participating in this joint effort, Infinigence AI is an AI infrastructure enterprise rooted in Xuhui's Model Space, committed to using an autonomous artificial intelligence infrastructure foundation to empower the intelligent revolution of all industries in the AI era; AgiBot is a leading enterprise in China's embodied intelligence track, with complete robot hardware and real test scenario capabilities. The two parties joined forces with Tsinghua University to carry out industry-university-research collaboration, bringing laboratory algorithms to real robot clusters and connecting the chain from technical research to industrial testing.

The value of the collective learning paradigm lies in turning the large-scale deployment of robots from an "endpoint of capability" into a "starting point of capability evolution." Robots no longer work in isolation but rely on collective experience for continuous iteration, alleviating the heavy dependence on manual demonstrations and providing a feasible path for humanoid robots to enter various scenarios such as homes and factories in the future. article_adlist[

Here's a news blind box for you

Come and open it and take a look ]article_adlistarticle_adlist[

Reporter: Wang Yongjuan

Editor: Zheng Liang

Please indicate the source when reprinting: Shanghai Xuhui Official Account ]article_adlistarticle_adlist[

]article_adlist news_keyword_pub,stock,sz300024 iframe below the main text iframe below the main text General statement General statement QR code at the end start QR code at the end start

QR code at the end start QR code at the end start Editor name and work code Editor name and work code