AgiBot Releases Embodied Agent Framework Genie Master, Validating Physical RSI in a Convenience Store Scenario

Zhiyuan Releases Embodied Agent Framework Genie Master, First to Validate Physical RSI in Real Application Scenarios

Zhong Lilei 4 minutes ago

From "how to get on the job" to "how to keep improving after getting on the job," AI WAVE further unfolds Zhiyuan's operational intelligence layout.

On October 27, as the fourth release during AI WAVE, Zhiyuan officially launched the embodied Agent framework Genie Master, enabling robots to "learn new skills while on the job." Genie Master allows robots to achieve autonomous improvement after deployment, continuously enhancing their capabilities in real work, thereby laying the foundation for cross-scenario replication and large-scale application.

When robots enter real scenarios, they will inevitably encounter many situations not seen during training. Relying on humans to chase after them to modify code and collect data means the deployment of embodied intelligence can only advance at a small, slow pace. How can robots "learn new skills while on the job"?

On October 27, Zhiyuan officially released the embodied Agent framework Genie Master to answer this industry question. The Jingling G2 equipped with this framework, during its deployment at a real convenience store, without relying on manual demonstration and debugging, locates the causes of failure itself, adjusts strategies, and accumulates experience, achieving continuous improvement on multiple complex and delicate tasks and realizing "human-out-of-the-loop" continuous self-evolution. This is the industry's first time stepping out of simulation environments and out of the laboratory, and the first validation of Physical RSI (embodied recursive self-improvement) in real business scenarios.

When robots can autonomously accumulate experience during deployment, the same set of capabilities can continuously grow thicker, and as they are replicated to more stores and more scenarios, marginal costs will continue to decline—this is precisely the key prerequisite for embodied intelligence to move toward scale. Building a bridge to support Physical RSI

When loading and unloading goods at a convenience store, the robot needs to hang products onto the hooks of the shelf. This seemingly simple task may encounter a series of problems: the hook size is too small, the label tag blocks the line of sight, and the head camera cannot clearly see the exact position. If the Agent only gives a single instruction to "hang the goods up," it is difficult for the action policy to obtain enough execution information; if the on-site results cannot be returned after a failure, it is difficult for the Agent to determine where the problem actually lies.

How can a robot, like a human, learn lessons from repeated practice and thus become more skilled the more it works?

The key is not only "being able to think," but also "being able to act, knowing the results, and continuing to improve." The judgments given by the Agent must be translated into the robot's specific actions; the real results after the robot executes must also return to the Agent and become the basis for the next round of judgment and learning.

Building a bridge that connects judgment, execution, and feedback is precisely the design starting point of Genie Master. To this end, the research team proposed the Embodied Execution Protocol EEP (Embodied Execution Protocol), which unifies four types of interfaces—capabilities, guidance, feedback, and continuity—so that the Agent clearly knows: which robot capabilities can be invoked, how this step should be guided for execution, whether the actual results meet the requirements, and what tasks remain to be completed.

With this protocol, the Agent issues execution guidance based on on-site observations; after the robot executes, information such as whether it succeeded and at which step it got stuck is returned in real time; once a task is interrupted, the system retains the progress already completed, and the Agent replans from the breakpoint rather than starting over from scratch. When multiple robots collaborate, the Agent can also continuously track the cluster's division of labor and task dependencies, supporting dynamic multi-robot scheduling.

In the aforementioned hanging task, after reviewing the failure records, the Agent found that the resolution of the head view was insufficient and the hook was blocked by the label tag, causing the robot to be unable to identify the hook position. Therefore, the Agent adjusted the execution guidance: the left hand first approaches the target and uses the wrist camera to supplement a close-range view; after confirming the hook position, the right hand completes the hanging. From discovering "it can't be hung up," to determining "why it can't be hung up," to adjusting "how it should be hung," the execution results thus become the basis for the next round of judgment.

Two routes teach robots to "learn from every fall"

Closing the execution loop is only the first step. To enable continuous improvement of robot capabilities, it is also necessary to turn a single lesson into long-term reusable experience.

The research team uses two complementary routes, Harness RSI and Policy RSI, to retain experience in prompts and memory, and in action policy parameters, respectively: one lets the robot remember experience, and the other lets the robot train experience into capability.

Harness RSI makes experience accumulable. During reasoning and execution, the robot records successful and failed cases as it works. After the daily tasks end, the model uniformly summarizes them and deposits the reusable content into long-term memory. When similar tasks and scenarios are encountered, this experience can be gradually loaded and invoked.

In the aforementioned hanging task, the adjustment process of "the head camera can't see clearly—switch to the wrist camera for positioning—the right hand completes the hanging" can be summarized into a reusable "action guide." The next time a similar problem is encountered, the Agent does not need to explore entirely from scratch but can invoke existing experience to organize its actions.

Policy RSI makes capabilities solidifiable. During the Agent's assistance in execution, takeover, and recovery, the system generates data containing problems, adjustments, and results. After post-processing, these data are used to train action policies, so that effective improvements not only serve the current task but are also retained in model parameters and transformed into operational capabilities that can be invoked later.

Taking takeout packing as an example, to address the problem that water bottles and grippers are easily entangled by white carrying straps, the Agent uses more specific action prompts to guide the gripper to retract first and then try again to place the item into the packaging bag. Such execution and recovery data will subsequently flow back into training and become material for subsequent action learning.

The two routes thus cooperate with each other: the former helps the Agent diagnose problems more effectively, while the latter transforms execution experience into retainable operational capabilities, thereby making the experience of this round of work the starting point for the next evolution. Entering a store, letting robots continuously evolve in real combat

The capabilities of Genie Master have already been preliminarily tested in real scenarios.

Over the past period, G2 has been officially on duty at a convenience store and has achieved continuous improvement on multiple complex and delicate tasks: prepared-food packing has become smoother after repeated failures, and the accuracy of picking and restocking actions has also steadily improved.

These changes do not come from manual parameter tuning but are the result of the robot locating problems and accumulating experience on its own during execution. This is also the industry's first time stepping out of simulation environments and out of the laboratory to validate Physical RSI in real scenarios—robots do not rely on manual demonstration and debugging and continuously self-evolve while deployed.

Validation in real business scenarios is only the first step. Next, the research team will continue to test and improve this system in more tasks and scenarios.

On November 2, Zhiyuan will launch a 24-hour uninterrupted livestream. Six Jingling G2 robots will serve as "probationary employees" of the robot convenience store, clocking in for work at 10:00 on the first day and completing a continuous 24-hour "confirmation review" by 10:00 the next day.

They will go through four shifts: the lunchtime peak trial shift, the evening peak stress-resistance shift, the late-night independent duty shift, and the early-morning breakfast sprint. The entire livestream is not just to show what actions the robots can complete, but more importantly to test whether they can be competent for work in a continuously changing real environment.

No going offline, no slacking off, able to serve, able to adapt. This time, can Jingling G2 pass the continuous 24-hour test and successfully "become a regular employee"?

Source: Submission

The images in this article are from: 123RF licensed image library. Infringement report Return to share Related articles All comments