Zhiyuan Robotics releases embodied Agent framework Genie Master, achieving Physical RSI in a convenience store

Physical RSI's first stop! The convenience store worker "person" evolves on its own, and the physical AI flywheel starts spinning

Because the small-sized hook on the shelf was partially blocked by the price tag, the robot-kun working at the convenience store tried to hang goods for the first time... and missed.

In the past, at this point a human engineer would have to step in, first troubleshoot the problem, then adjust the program and supplement the data so the robot could practice again.

But robot-kun is a little different. It figured out a new approach on its own.

It reviewed its previous execution records and found that the resolution of the head camera's field of view was insufficient, and combined with the occlusion by the tag, it was difficult to see the position of the small hook clearly.

So the Agent adjusted its strategy, having the left hand approach the target first and use the wrist camera to add a close-up view; after confirming the position, the right hand completed hanging the goods.

More notably, robot-kun not only discovered the problem and found a solution on its own just now, but also quietly accumulated this effective experience for future use.

The mechanism that lets the robot discover problems, adjust strategies, and accumulate experience during work is backed by AgiBot's newly released embodied Agent framework Genie Master.

This is also the fourth installment of AgiBot's recently launched "AI WAVE" series of technology releases—this time, AgiBot has brought Physical RSI (embodied recursive self-improvement) into a real convenience store business flow, a first in the industry.

With the support of Genie Master, AgiBot's Genie G2 robot, while executing tasks in a real convenience store environment, can already analyze the causes of failure, adjust strategies, and accumulate experience autonomously without relying on human demonstrations and debugging, achieving continuous self-improvement with "no human in the loop."

This is the first time in the industry that Physical RSI has been run through and validated in a real business scenario, breaking out of simulation environments and laboratory toy settings.

How exactly does Genie Master's self-evolving closed loop work? And what disruptive changes can it bring to the large-scale deployment of robots?

What is Genie Master?

To understand why this robot can "enlighten itself," we first need to look at what Genie Master, the framework behind it, actually is.

Simply put, Genie Master is an embodied Agent framework whose core task is to connect the reasoning capabilities of the upper-level Agent with the action execution capabilities of the lower-level robot.

It not only enables the Agent to smoothly invoke the robot's various capabilities and understand real physical execution feedback, but also allows it to continuously organize actions and targeted practice based on task progress.

Once the robot exposes a capability shortfall in actual work, the problem goes directly into the autonomous improvement process.

It is not hard to see that Genie Master aims to build a continuous improvement mechanism after robot deployment—this involves both judgment and dynamic adjustment at the moment of task execution, as well as post-hoc experience accumulation and updates to the underlying action policies.

AgiBot proposed this framework because it recognized a long-standing pain point in the industry's deployment.

Industry insiders know that once robots enter real commercial scenarios, they inevitably encounter long-tail situations not covered during training.

Execution failures, unstable actions, or a task taking too long may all require engineers to intervene and troubleshoot, then improve performance by adjusting programs, supplementing demonstration data, or training models.

The more scenarios there are and the more complex the tasks, the more R&D and maintenance workload increases.

If every new requirement waits for humans to troubleshoot and debug from scratch, iteration speed will be limited, making it even harder to scale deployment across scenarios.

In a word, for embodied intelligence to be truly used at scale, it must shift from "passive fine-tuning" that relies on human demonstrations to a "self-evolution mode" in which robots autonomously discover and solve problems.

Genie Master hopes to hand part of this work over to the Agent.

By putting problem diagnosis, organizing practice, and result evaluation all under the Agent's control, Genie Master supports robots in autonomously acquiring new capabilities, thereby greatly shortening the R&D cycle from "a customer raises a new requirement" to "the robot masters this new skill."

In addition to the hook restocking mentioned at the beginning, in real convenience store tests, the robot equipped with Genie Master also demonstrated more capabilities for handling complex situations.

According to AgiBot, the Genie Master team has incorporated picking, restocking, cleaning, exception recovery, and multi-robot collaboration into a unified workflow, exploring a continuous improvement loop for robots in real business operations.

Practicing on its own during idle time: the embodied Agent framework connects the brain and the cerebellum

Getting robots to self-evolve during work is not a difficult idea.

But before this, why did this matter remain only in laboratory discussions, with almost no one actually doing it and making it happen in real business?

The reason is that both the accumulation of experience at the upper level and the internalization of policies at the lower level must be built on an extremely hardcore foundation—

The guidance of the upper-level reasoning "brain" must be precisely translated into actions; the results of the lower-level execution "cerebellum" must also be fed back to the brain in real time.

To connect this vital pathway, Genie Master has introduced one of its key designs:

The Embodied Execution Protocol EEP (Embodied Execution Protocol).

EEP establishes a unified interface between the Agent and the physical world, mainly covering four core types of interfaces:

Capability interface: lets the Agent clearly know what capabilities the robot currently has available to call;

Guidance interface: enables the Agent to pass specific spatial hints, object references, and fine-grained language requirements to the low-level action execution end;

Feedback interface: lets the Agent understand in real time the actual physical situation after the robot executes an action;

Continuity interface: enables the system to continuously track the task progress and division-of-labor dependencies of the robot cluster, supporting dynamic scheduling and task resumption after interruptions.

Through the EEP protocol, together with collaborative designs such as Temporal Alignment, the upper-level Agent can dynamically adjust guidance at any time based on execution results; while the low-level Action Policy can precisely execute actions by combining current observations with the Agent's requirements, and return observations and subtask status in real time.

In an embodied intelligence system, only by truly twisting reasoning, execution, and feedback into a single rope can self-evolution cease to be a castle in the air.

EEP plays exactly the key interface role in this interaction chain, laying the foundation for subsequent experience accumulation and policy updates.

Once the foundation is solid, the robot's daily "self-grinding" in the deployment state begins.

In the normal workflow of a convenience store, when the robot finds that certain tasks repeatedly fail, execute unstably, or take too long—that is, when "capability gaps" are exposed—the Agent will not sit idly by.

It will proactively use the convenience store's idle time to organize the robot to conduct "autonomous extra practice" targeting these weak links.

What to practice, how to practice, whether the results after practice are qualified, and how to recover and retry when anomalies occur—all are arranged single-handedly by the Agent. The real deployment environment thus continuously provides the robot with real-world evidence for "what to practice next and what to change next."

One route accumulates experience, one route updates the action policy

Now that "discovering problems" and "autonomous extra practice" have been arranged in real scenarios, how does the robot truly grow the results of its practice into its own abilities?

Here, Genie Master adopts a "two-pronged" strategy.

One route is called Harness RSI, responsible for accumulating experience; the other route is called Policy RSI, responsible for updating the action policy.

Let's look at Harness RSI first.

Its core logic is, while keeping model weights and existing tool interfaces unchanged, mainly improving subsequent task performance by optimizing prompts, memory, and execution rules.

During daily reasoning and execution, the robot "records as it works" both successful and failed examples; then, through daily asynchronous reflection and memory consolidation, it transforms these execution records into more effective guidance, completion determination, anomaly recovery, and collaboration rules, incorporating them into long-term memory like writing an "error notebook," and gradually loading them as needed by task and scenario later.

Looking back at the "convenience store hanging goods" case at the beginning of the article, it is actually Harness RSI at work.

At first, the success rate of that small hook task was very low. But after reviewing the failure records, the Agent diagnosed the root cause, namely "insufficient resolution of the head camera view, plus occlusion by the label tag, which prevented the robot from accurately identifying the position of the small hook."

So the Agent wrote the new solution of "first use the left wrist camera to move closer and supplement the view, and after confirmation use the right hand to hang the goods" into executable experience; when similar situations occur later, it can directly retrieve the experience and handle them calmly.

Beyond the "error notebook," in order to meet the strict requirements of real-machine deployment for low latency and high stability, some capabilities must become the robot's "muscle memory."

This is where Policy RSI comes in.

Policy RSI collects and organizes all the data generated by the robot during autonomous execution, manual takeover demonstrations, and anomaly recovery, improves data quality through filtering, weighting, and target relabeling, and then uses supervised learning, imitation learning, or reinforcement learning to directly update the low-level Action Policy model, so that improvements are truly retained in the model parameters.

Suppose the robot is performing the task of "preventing tangling during takeout packaging" in the store; the data generated when handling strap tangling, executing retraction, and recovering from anomalies can, after filtering and processing, be used for subsequent action policy training to help the model improve execution performance on similar tasks.

Through Policy RSI, the robot can absorb nutrients from real execution experiences again and again, gaining more stable execution capability and more accurate instruction-following ability.

It should be noted that this does not mean human participation is no longer needed at all; Policy RSI still uses data such as manual takeover demonstrations, and updating the action policy also requires data processing and model training.

What's even better is that these two routes do not fight separately, but interlock and form a flywheel that spins faster and faster—

Harness RSI is responsible for accumulating and organizing experience, helping the system analyze failure causes, choose practice methods, and improve subsequent execution;

Policy RSI uses processed execution data for action policy training, giving effective experience a chance to settle into more stable action capabilities.

The two routes promote each other, forming a closed loop from real business execution, discovering capability gaps, autonomous practice, to experience accumulation and policy updates. AgiBot hopes that through this, the experience accumulated in each round can help the system more effectively diagnose problems, choose practice, and utilize data.

Let the ability to improve capabilities also continuously evolve, and obtain greater capability increments under the same investment—this is exactly where the true power of "Recursive Self-Improvement (RSI)" lies.

From "fixed at factory" to "the more it works, the more capable it becomes"

Stepping outside the convenience store and looking from the perspective of the entire embodied intelligence industry, the direction explored by AgiBot Genie Master is pushing open a new door.

First, in terms of R&D efficiency, it completely breaks the shackles of the previous "fixed at factory" robot capabilities and shifts them to a new paradigm of "continuous growth during deployment."

Second, in terms of commercialization, this also settles an economic account clearly.

When robots "become more capable the more they work" in real operations, the same set of capabilities will continuously thicken during deployment.

When replicated to more stores and more cross-industry scenarios in the future, the marginal cost of subsequent deployment and R&D will continue to decline.

And this continuously declining cost curve is precisely the core prerequisite for embodied intelligence to move toward large-scale commercialization.

Looking back at the "AI WAVE" series of releases that AgiBot launched starting October 22, as the fourth item in the series, Genie Master plays a key connecting role in AgiBot's overall full-stack embodied AI layout—it extends and roots the underlying foundation models and full-stack AI capabilities into the real commercial deployment state.

Of course, no matter how brilliantly the technical architecture is described, ultimately it still has to be judged by real-world performance.

As the saying goes, whether it's a mule or a horse, take it out for a walk.

On November 2, AgiBot plans to launch a "24-hour convenience store livestream" watched by the whole internet.

At that time, 6 Lingxi G2 robots will take up posts continuously, and in real convenience store operations without a script, they will undergo a strict probation assessment.