AnswersPublished16 sources
Collective Experience Learning: How Robots Learn From a Fleet’s Real-World Experience
Collective Experience Learning (CEL) lets deployed robots improve a shared policy using fleet experience instead of relying on new expert demonstrations for every update. FERT uses failure contexts to retrieve similar successful segments for training, while FATS supports fault tolerant asynchronous fleet learning.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Collective Experience Learning (CEL) lets deployed robots improve a shared policy using fleet experience instead of relying on new expert demonstrations for every update.
- FERT uses failure contexts to retrieve similar successful segments for training, while FATS supports fault tolerant asynchronous fleet learning.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is Collective Experience Learning (CEL), proposed by Tsinghua University, Infinigence AI and AgiBot and described in an Infinigence AI–. Article summary: Collective Experience Learning (CEL) is a fleet-learning approach in which deployed robots share real-world execution experience to improve a common policy, rather than relying on fresh human demonstrations for every imp. Topic tags: general, academic, education, general web, documentation. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, waterm
Collective Experience Learning (CEL) is a way for deployed robots to improve a shared policy by learning from experience gathered across a fleet. Rather than depending on people to demonstrate or correct every new behavior, the approach uses successful robot runs as training material. A study associated with Tsinghua University, Infinigence AI and AgiBot reports tests across a fleet of up to 96 robots. 1
9
Why share robot experience?
Human demonstrations and corrective supervision can help robots learn, but providing them repeatedly becomes difficult as deployments grow. CEL shifts some of that learning burden to the robots themselves: a useful execution by one robot can become an example for the wider fleet. The goal is not to eliminate human involvement in robotics, but to make improvement less dependent on fresh expert input for every update. 1
How FERT turns failures into learning signals
CEL’s central algorithm is Failure-guided Experience Retrieval Fine-tuning (FERT). It uses the context around a failed execution to search a shared collection of successful trajectories for relevant segments. The successful segment supplies recorded actions for policy updates; the failure context helps identify which experience may be useful. 4
13
This is particularly relevant to long manipulation tasks. A final success-or-failure label says whether a run worked, but does not by itself identify which step needs improvement. FERT offers a way to connect a problem context with an example of a successful approach. It does not, by itself, prove which exact action caused a failure.
The team also describes FATS, a fault-tolerant asynchronous training system intended to support learning across robots that may not all be available at the same time. Reporting on the system cites 30 faults affecting 22 robots, with 1.3% of robot operating time lost. These figures describe reported system performance; they do not independently establish every detail of its fault-recovery or upload mechanisms. 8
What the reported evaluations show
The evaluation covered five manipulation tasks: flower arranging, clothes folding, phone charging, stool assembly and phone packaging.
- Performance on held-out robots: Two robots that did not participate in training completed 600 trials, with 585 successes and 15 failures—a reported overall success rate of 97.5%. The team also reports that the robots’ results on each task exceeded the average of six human operators. This is a comparison within the reported tasks, not evidence of general superiority to people.
- Phone packaging: Reported success increased from 79.2% to 98.3%. Coverage of the evaluation describes 120 tests on two robots that had not participated in training.
7
- Fleet size: At equal real-world training time, a reported comparison increased success from 86.7% with 16 robots to 98.3% with 96 robots. This supports the idea that a larger fleet can contribute more experience within the same elapsed training time, though it is not a guarantee that every larger fleet will produce the same gains.
What to take away—and what remains uncertain
CEL’s main idea is straightforward: use failures to identify where improvement is needed, then draw on successful examples collected by the fleet. FERT provides that retrieval-and-update mechanism, while FATS is described as the system layer for asynchronous, fault-tolerant training. The reported results are promising early evidence across five real-world tasks, including tests on robots held out from training. 13
The performance figures should still be treated as team-reported results, not independent replication. They show how CEL performed in the reported setup; they do not establish how well it will transfer to different robots, tasks or deployment conditions.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Collective Experience Learning: How Robots Learn From a Fleet’s Real-World Experience"?
Collective Experience Learning (CEL) lets deployed robots improve a shared policy using fleet experience instead of relying on new expert demonstrations for every update.
What are the key points to validate first?
Collective Experience Learning (CEL) lets deployed robots improve a shared policy using fleet experience instead of relying on new expert demonstrations for every update. FERT uses failure contexts to retrieve similar successful segments for training, while FATS supports fault tolerant asynchronous fleet learning.