AgiBot's Genie Master lets store robots diagnose and fix their own failures

Chinese robotics company AgiBot has detailed Genie Master, an embodied agent framework that lets its G2 robots diagnose failures, adjust strategies and bank the resulting lessons without an engineer in the loop. It is the fourth release in the company's AI WAVE series, which began on October 22.

The illustrative case is deliberately mundane. A small hook on a convenience-store shelf was partly hidden by a price tag, and the robot missed on its first attempt to hang an item. Instead of waiting for a human to debug the program, it reviewed its own execution logs, concluded that its head camera lacked the resolution to locate the hook through the obstruction, and re-planned: bring the left hand in close with a wrist camera to confirm position, then hang the item with the right hand. It stored the fix for later use.

Genie Master sits between an upper-level agent's reasoning and the robot's low-level motion. It lets the agent call robot capabilities, read physical execution feedback, and organize further practice based on how each task is progressing.

The connective tissue is the Embodied Execution Protocol, or EEP, which defines capability, guidance, feedback and continuity interfaces. Paired with temporal alignment work, it translates an agent's intent into actions the action policy can execute and returns observations and sub-task state in real time.

Improvement runs on two tracks. Harness RSI leaves model weights untouched and rewrites prompts, memory and execution rules, accumulating successes and failures into something like an error notebook. Policy RSI collects data from autonomous runs, human takeovers and recovery attempts, filters and reweights it, then updates the underlying action policy through supervised, imitation or reinforcement learning.

The two tracks feed each other. Real deployments surface capability gaps, the agent schedules its own practice during idle hours, and the resulting experience either becomes retrievable guidance or gets baked into the policy itself. AgiBot says picking, restocking, cleaning, anomaly recovery and multi-robot coordination now sit in one workflow.

No human in the loop is not the same as no humans. Policy RSI still draws on takeover demonstrations, and updating the action policy still requires data processing and model training.

The commercial argument is about marginal cost: if competence keeps growing after deployment rather than being fixed at the factory, the cycle from a customer request to a working skill shortens. The public test comes on November 2, when six G2 robots are scheduled to work a real, unscripted convenience store on a 24-hour livestream.