正文 · AI 翻译
An Unassisted In-Store Test Tackles Long-Tail Scenarios
By Hengyu, reporting from Aofei Temple
QbitAI | WeChat public account QbitAI
By Hengyu, reporting from Aofei Temple
QbitAI | WeChat public account QbitAI
Because a price label partly obscured a small hook on a shelf, the robot working in a convenience store missed it on its first attempt to hang a product.
In the past, a human engineer would have had to step in, diagnose the problem, adjust the program, add data, and have the robot practice again.
But this robot was a little different. It worked out a new approach on its own.
It reviewed its earlier execution records and found that the head camera's view lacked sufficient resolution and, with the label in the way, it was hard to see where the small hook was.
The Agent therefore changed its strategy: it moved the left hand toward the target first, used the wrist camera for a close-up view, confirmed the position, and then used the right hand to hang the product.
What is more, the robot had just identified the problem and found a solution on its own, then quietly retained that useful experience for future use.
Behind this mechanism for finding problems, adjusting strategies, and accumulating experience while working is Genie Master, the embodied Agent framework recently released by AgiBot.
This is also the fourth installment in AgiBot's recent “AI WAVE” series of technology releases. This time, AgiBot has brought Physical RSI (embodied recursive self-improvement) into a real convenience-store workflow, which it describes as an industry first.
With Genie Master, AgiBot's Elf G2 robot can now analyze the causes of failures, adjust its strategy, and retain experience while carrying out tasks in a real convenience store, without relying on human demonstrations or debugging, achieving continuous self-improvement with “no human in the loop.”
This is the industry's first demonstration and validation of Physical RSI in a real business setting beyond simulations and laboratory toy settings.
How does Genie Master's autonomous improvement loop work? And what disruptive changes could it bring to deploying robots at scale?
What is Genie Master?
To understand why this robot can “figure things out for itself,” we first need to look at what Genie Master, the framework behind it, actually is.
Put simply, Genie Master is an embodied Agent framework whose core task is to connect the reasoning capabilities of a higher-level Agent with the action execution capabilities of the robot beneath it.
It enables the Agent to call on the robot's capabilities and understand feedback from real physical execution, while continuously organizing actions and targeted practice as the task progresses.
Once the robot reveals a capability gap during actual work, that problem goes directly into an autonomous improvement process.
Genie Master is clearly focused on building a mechanism for continuous improvement after a robot is deployed. This covers both decisions and dynamic adjustments during task execution and the subsequent retention of experience and updating of underlying action policies.
AgiBot proposed this framework because it saw a longstanding obstacle to real-world deployment across the industry.
Industry insiders know that robots entering real commercial settings inevitably encounter long-tail situations that training did not cover.
Execution failures, unstable movements, or tasks that take too long can all require an engineer to investigate, then improve performance by adjusting programs, adding demonstration data, or training models.
As settings multiply and tasks grow more complex, the work involved in development and maintenance also increases.
If every new requirement has to wait for a person to investigate and debug it from scratch, iteration slows and large-scale deployment across settings becomes harder.
In short, for embodied intelligence to be used at scale, it must move from “passive fine-tuning” that depends on human demonstrations to a “self-improvement model” in which robots discover and solve problems on their own.
Genie Master aims to hand some of that work to the Agent.
By putting the Agent in charge of diagnosing problems, organizing practice, and evaluating results, Genie Master supports robots in learning new capabilities on their own, substantially shortening the development cycle between “a customer requests a new capability” and “the robot masters it.”
Beyond restocking products on hooks, as described at the start, robots equipped with Genie Master showed further abilities to handle complex situations in tests at a real convenience store.
According to AgiBot, the Genie Master team has brought product retrieval, restocking, cleaning, recovery from anomalies, and multi-robot collaboration into a unified workflow to explore a continuous improvement loop in real business operations.
Practicing on Its Own During Downtime: An Embodied Agent Framework Connects the Brain and Cerebellum
The idea of enabling robots to improve themselves while working is straightforward.
But why, until now, has it largely remained a topic of laboratory discussion, with almost no one making it work in real business operations?
The reason is that both the retention of experience at the higher level and the internalization of policies at the lower level depend on an extremely demanding foundation—
Guidance from the reasoning “brain” at the higher level must translate precisely into actions, and the results produced by the execution “cerebellum” at the lower level must be fed back to the brain in real time.
To establish this vital connection, Genie Master uses one of its key designs:
The Embodied Execution Protocol (EEP).
EEP establishes a unified interface between the Agent and the physical world, mainly covering four types of core interface:
Capability interface: lets the Agent know which robot capabilities are currently available to call on;
Guidance interface: lets the Agent pass specific spatial cues, object references, and fine-grained language instructions to the underlying action executor;
Feedback interface: lets the Agent understand the actual physical situation in real time after the robot executes an action;
Continuity interface: lets the system continuously track task progress and dependencies in the division of work across a robot fleet, supporting dynamic scheduling and the resumption of tasks after interruptions.
Through EEP, together with coordinated designs such as Temporal Alignment, the higher-level Agent can adjust its guidance dynamically at any time based on execution results. The underlying foundation Action Policy can then execute actions precisely using current observations and the Agent's requirements, and return observations and subtask status in real time.
In an embodied intelligence system, self-improvement becomes feasible only when reasoning, execution, and feedback truly work together.
EEP provides the key interface in this interaction chain, laying the groundwork for subsequent accumulation of experience and policy updates.
Once that foundation is in place, the robot's routine of pushing itself to improve while deployed begins.
In the convenience store's normal workflow, when the robot finds that certain tasks repeatedly fail, are executed inconsistently, or take too long—revealing a “capability gap”—the Agent does not simply leave it alone.
It proactively uses the convenience store's downtime to organize “independent extra practice” for the robot on those weak points.
The Agent takes charge of what to practice, how to practice, whether the result meets the standard, and how to recover and retry when something goes wrong. In this way, the real deployment environment continually gives the robot practical evidence of “what to practice next and what to improve.”
One Path Accumulates Experience; the Other Updates Action Policies
Having arranged for the robot to “find problems” and “practice independently” in a real setting, how does it turn the results of that practice into capabilities it actually possesses?
Here, Genie Master takes a two-pronged approach.
One path, Harness RSI, accumulates experience; the other, Policy RSI, updates action policies.
First, consider Harness RSI.
Its core approach is to improve performance on later tasks mainly by optimizing prompts, memory, and execution rules, while keeping model weights and existing tool interfaces unchanged.
During routine reasoning and execution, the robot records examples of success and failure as it works. Then, through daily asynchronous reflection and memory consolidation, it turns those execution records into more effective guidance, completion criteria, anomaly recovery procedures, and collaboration rules. Like entries in a notebook of past mistakes, these become part of long-term memory and are gradually loaded later as tasks and settings require.
The convenience-store product-hanging example at the start of the article shows Harness RSI in action.
At first, the success rate for the small-hook task was very low. But after reviewing failure records, the Agent diagnosed the root cause: “the head camera's view lacked sufficient resolution, and the label obstructed it, preventing the robot from accurately identifying the position of the small hook.”
The Agent then wrote the new solution—“first move the left wrist camera closer to get another view, then use the right hand to hang the product after confirming the position”—into executable experience. When a similar situation arises later, it can retrieve that experience and handle it confidently.
Beyond this “notebook of past mistakes,” some capabilities must become the robot's “muscle memory” to meet the stringent low-latency and high-reliability requirements of deployment on physical robots.
That is where Policy RSI comes in.
Policy RSI collects and organizes data generated when the robot acts autonomously, during demonstrations involving human takeover, and during recovery from anomalies. It improves data quality through selection, weighting, and goal relabeling, then uses supervised learning, imitation learning, or reinforcement learning to directly update the underlying Action Policy model, retaining the improvements in the model's parameters.
Suppose a robot in the store is doing a “preventing tangles while packing takeout orders” task. Data produced as it handles tangled carrying straps, retracts its movements, and recovers from anomalies can, after selection and processing, be used to train later action policies and help the model perform similar tasks better.
Through Policy RSI, the robot can learn from repeated real-world execution, gaining more stable execution and a more accurate ability to follow instructions.
This does not mean human participation is no longer needed at all: Policy RSI still uses data such as demonstrations involving human takeover, and updating action policies still requires data processing and model training.
Better still, these two paths do not operate separately. They fit together to form a flywheel that spins ever faster—
Harness RSI accumulates and organizes experience, helping the system analyze the causes of failures, choose how to practice, and improve subsequent execution;
Policy RSI uses processed execution data to train action policies, giving effective experience a chance to become more stable action capabilities.
The two paths reinforce each other, forming a loop from real business execution and identification of capability gaps to independent practice, retention of experience, and policy updates. AgiBot hopes that the experience accumulated in each cycle will help the system diagnose problems, choose practice, and use data more effectively.
Continuously improving the ability to improve capabilities, and gaining more capability from the same investment: that is the real power of “recursive self-improvement (RSI).”
From “Fixed at the Factory” to “More Capable With Every Job”
Looking beyond the convenience store to the embodied intelligence industry as a whole, the direction AgiBot is exploring with Genie Master is opening a new door.
First, in development efficiency, it breaks the old constraint that a robot's capabilities are “fixed at the factory,” shifting them toward a new model of “continuous growth during deployment.”
When AI's new solutions can become new robot skills on site without outside direction, the delivery cycle from a customer's new request to a robot mastering the new skill will be greatly shortened.
Second, for commercial deployment, the economics also become clearer.
When robots become more capable as they work in real operations, the same set of capabilities will keep growing during deployment.
As this is replicated across more stores and more settings in different industries in the future, the marginal costs of subsequent deployment and development will continue to fall.
That steadily declining cost curve is a core prerequisite for embodied intelligence to reach large-scale commercial deployment.
Looking back at Zhiyuan’s “AI WAVE” series of announcements, which began on October 22, Genie Master is the fourth installment and plays a key role in connecting the stages of Zhiyuan’s overall full-stack embodied AI strategy: it extends and embeds the underlying foundation model and full-stack AI capabilities in real commercial deployments.
Of course, however impressive the technical architecture sounds, what ultimately matters is how it performs in practice.
As the saying goes, the proof is in the pudding.
On November 2, Zhiyuan plans to launch a “24-hour convenience store livestream” for viewers across the internet.
At that time, 6 Genie G2 robots will work continuously, undergoing a rigorous assessment for permanent employment during the unscripted operation of a real convenience store.