AgiBot proposes Worken metric to measure robot work with the goal of reducing the engineer-to-robot ratio to 1:10 by year-end

Researchers from Tsinghua University, Infinigence AI, and AgiBot have introduced a learning approach called Collective Experience Learning (CEL). This method allows robots to learn from the successes of other robots instead of relying primarily on human demonstrations.

According to a report published by QbitAI on October 30th, the team tested this methodology on a cluster of 96 real robots. The corresponding scientific paper is available on arXiv under the number 2610.25736.

The original problem that the researchers are solving relates to labor resources. Robots improve by accumulating practical experience, but today this is most often achieved through repeated movements by humans or operator intervention to correct errors. Although hundreds of robots can be deployed, qualified operators cannot be scaled at the same pace.

When using CEL, successful experience gained by one robot is saved and shared, allowing a robot facing a task to learn how its colleagues resolved similar situations.

Task Testing and Results

The team tested five different tasks covering both domestic and factory work: sorting colors, folding clothes, charging a phone, assembling a stool, and packaging a phone. Two robots not involved in the training performed 600 trials, incurring 15 failures, which resulted in an overall success rate of 97.5 percent. The success rate across all five tasks exceeded the average performance of six human operators.

Specifically, in phone packaging, which is a multi-stage industrial task, the success rate increased from 79.2 percent to 98.3 percent after applying collective learning, with only two failures recorded in 120 consecutive attempts.

A scaling trend was also noted: while keeping the same amount of real-time training, increasing the number of participating robots from 16 to 96 raised the success rate from 86.7 percent to 98.3 percent.

Technical Components of the System

The operation of this system is ensured by two key elements. The first is a fault-tolerant asynchronous learning system named FATS. This system is built on the open-source large-scale reinforcement learning platform RLinf. Robots upload their experience immediately after completing a task, without waiting for others, and the cluster continues to operate even if devices disconnect.

During the test with 96 robots, the cluster collected 9,834 episodes. Twenty-two robots had a total of 30 failures, but the loss of robot time was only 1.3 percent of the total operating time.

The second component is an algorithm called failure-guided experience retrieval fine-tuning (FERT). One manipulation can last minutes and involve thousands of action steps, but feedback often boils down only to final success or failure. FERT uses both aspects: when a robot encounters difficulties, it retrieves successful experience from analogous situations and focuses the learning specifically on these examples.

The report was provided by Infinigence AI. The results are presented by the research group and have not yet undergone independent verification.

AgiBot has introduced Genie Master—an embodied agent framework that enables robots to continue improving after deployment, rather than waiting for engineers to conduct retraining. The company states that G2 robots using this framework operated in a real neighborhood store, performing tasks such as packaging ready meals, picking items, and restocking shelves, while also detecting and resolving their own failure modes without human intervention in the training process.

AgiBot describes the result as the first proof of concept for what they call Physical RSI, or Embodied Recursive Self-Improvement, outside of simulation and laboratory settings. One example presented at the presentation involved a robot repeatedly missing a small hook on a shelf because a price tag was obstructing the view from its head camera. By analyzing its own execution logs, the agent identified the problem, changed its approach by bringing its left arm closer so the wrist camera could confirm the hook's position, then picked up the item with its right hand and saved this correction for future use.

This framework is based on the Embodied Execution Protocol (EEP)—a unified interface between the agent and the robot. It consists of four components: a capabilities interface, which informs the agent what the robot can do; a guidance interface for transmitting spatial cues, object references, and detailed action policies; a feedback interface, which reports the physical outcome; and a continuity interface, which tracks progress and task dependencies across the entire robot fleet, allowing work to resume or be rescheduled after breaks.

When tasks consistently fail or are performed slowly, the agent schedules targeted practice during the store's least busy hours. Improvement occurs in two directions. RSI, which does not affect model weights, refines cues, memory, and execution rules by using daily analysis of successes and failures to create what the team compares to an error log. The RSI policy collects data from autonomous runs, human interventions, and recoveries, filters and reweights it, and then retrains the base action policy using supervised learning, imitation, or reinforcement. AgiBot notes that the second track still utilizes human intervention data and model training, meaning humans are not entirely excluded.

Genie Master is the fourth release in AgiBot's AI WAVE series, which began on October 22nd. According to the company's published schedule, the next item on October 28th will be an update to the Genie Sim simulation platform with an internal physics engine designed to simulate deformable objects, which, according to AgiBot, achieves over 95% consistency between real and simulated results in complex tasks. More detailed information was not published at the time of writing.

Public testing is scheduled for November 2nd and 3rd, when six G2 robots will operate for 24 hours in a real neighborhood store environment, covering peak lunch and dinner hours, the night shift, and breakfast preparation.

Despite the current dominance of the artificial intelligence (AI) era, which allows for the instant creation of articles, speeches, and emails, human intellect remains indispensable for developing true strategy, understanding, and vision. A recent decision by Anthropic serves as a clear confirmation of this.

Anthropic, the developer of the popular chatbot Claude, is currently recruiting an experienced writer. The company is prepared to offer a high annual salary of 3.25 crore rupees for this position (equivalent to approximately $295,000 to $345,000 USD). According to the official job posting, this position is available in San Francisco or New York.

The role is not limited to simple text writing; the selected candidate will collaborate closely with senior management. They will be responsible for explaining complex technical and developmental concepts in simple language, preparing communication materials for major events, and defining how the company should present itself to the world.

The primary duties of this writer include working with Anthropic's senior executives, the external communications team, and subject matter experts. Tasks involve preparing and editing executive communications, preparing leaders for public appearances and interviews, and assisting in rapid and accurate responses to any breaking news.

Furthermore, the company requires a specialist who has a deep understanding of advanced AI technologies and can explain these complex topics to a wide audience, whether they are experts or general users.

Anthropic is looking for established professionals, not graduates. Official requirements include at least eight years of experience in a large organization involved in high-level corporate communications. Outstanding writing skills and the ability to perform under pressure, as well as the capacity to make balanced decisions and maintain a clear strategic stance amid uncertainty, are also necessary.

The minimum required qualification is a bachelor's degree or equivalent professional experience. Candidates with a good understanding of science and research, AI products and architecture, corporate strategy, visual storytelling, and brand communication are of particular interest.

Candidates meeting these criteria can apply through Anthropic's official career portal. This example demonstrates that even if machines have learned to write words, the art of understanding complex ideas and determining what and when to say remains a human prerogative, and major technology companies like Anthropic are willing to pay a high price for this skill.