AgiBot has introduced Genie Master, a software framework designed to help robots improve their performance while carrying out real-world tasks. The Chinese humanoid robotics company demonstrated the system on its wheeled G2 robot in a convenience store, where the robot used feedback from unsuccessful attempts to adjust how it handled products and completed stocking tasks.
Announced on October 27 as part of AgiBot’s AI WAVE series, Genie Master connects an AI agent’s planning decisions with the robot’s physical actions and the feedback generated during execution. The company calls the approach Physical RSI, or embodied recursive self-improvement, and says it allows robots to learn from operational mistakes without requiring an engineer to manually demonstrate each correction.
At the center of the framework is an Embodied Execution Protocol, or EEP, which organizes four kinds of information: the robot’s available capabilities, instructions for carrying out an action, feedback on the result and the progress of unfinished tasks. If an operation fails, the agent can use that information to identify what went wrong and revise its approach rather than restart the entire sequence.
In one example, a G2 robot struggled to hang merchandise because a label obstructed the hook and its head-mounted camera could not see the target clearly. The agent changed the sequence, moving the robot’s left hand closer to obtain a better view through its wrist camera before using the right hand to finish the placement.
AgiBot describes two ways of preserving these corrections. Harness RSI records successful and unsuccessful attempts and consolidates reusable lessons into long-term memory. Policy RSI uses execution and recovery data to further train the robot’s movement policies, potentially making improvements available during subsequent tasks.
The company says its G2 robots have shown more reliable performance in convenience-store activities including food packaging, retrieving products and replenishing shelves. It has not published independently verified figures for task success rates, improvement over time or the number of human interventions required.
AgiBot plans to put six G2 robots through a publicly streamed 24-hour convenience-store shift from November 2 to 3, covering stocking, food preparation, cleaning and customer-facing activities across four work periods.
