AGIBOT is trying to cut the engineering work each new robot site needs by letting robots learn on site, a claim its November 2 livestream will test in public.
AGIBOT has released Genie Master, a software framework that lets its robots improve at their jobs after they are deployed, and has tested it on G2 robots working in a real convenience store. IT Home reported the October 27 launch. AGIBOT says Genie Master lets its robots find the cause of a failure, adjust their approach and store the lesson without a human demonstrating the task again.
AGIBOT describes the trial as the first time robot self-improvement of this kind, which it calls Physical RSI, for recursive self-improvement, has been validated outside simulation in a real business setting, according to PEdaily. That is the company’s own claim. The tasks were ordinary store work: packing hot food, picking items and restocking shelves.
AGIBOT has not published improvement rates, named the store or released code or a paper. It plans a public test instead: six G2 robots will run a convenience store for 24 hours without a break, starting on November 2.
Two learning loops do different jobs
The framework rests on what AGIBOT calls the Embodied Execution Protocol, a common interface that tells the controlling AI agent what the robot can do, how its actions turned out physically and where each robot in a group stands on a shared task, QbitAI reported. When a robot fails partway through hanging an item, the system keeps the progress it has already made rather than starting over.
Improvement then runs on two tracks. The first, which AGIBOT calls Harness RSI, changes the instructions, memory and execution rules around the robot’s AI model while leaving the model itself untouched, through daily reflection and memory consolidation run in the background. The second, Policy RSI, takes data from robot runs and human corrections and retrains the underlying action model, so that gains are stored in the model’s parameters. QbitAI described the two together as a flywheel that speeds up as it turns.
In one example QbitAI gave, a robot’s first attempt to hang goods on a small shelf hook missed because a price tag partly hid the hook from its head camera. The agent wrote a new step into its stored experience: bring the left wrist camera in for a closer view, then hang the item with the right hand. Neither QbitAI nor IT Home gave success rates before or after the change.
A week of releases leads to Innovation Day
Genie Master is the fourth release in AGIBOT’s "AI WAVE" series, which began on October 22 with an open dataset for physical reasoning, followed by its Genie Studio 2.0 software and, on October 26, the G2 Max heavy-lift robot, the Yangtse Evening Post reported. The schedule lists an upgraded Genie Sim simulator on October 28, a foundation model on October 29 and a system for robots to learn from each other’s experience on October 30. AGIBOT holds its Innovation Day on October 29.
On the same day as Genie Master, AGIBOT said a preview of its GO-3 Nano model ranked first on the RoboDojo leaderboard, with an average score of 61.90 and an average success rate of 56.42%, ahead of OpenAI’s GPT-6 Astra, Jiemian reported. AGIBOT disclosed the result itself. RoboDojo is a 42-task test run on real dual-arm robots, which a SenseNova-led study used to evaluate GPT-6 Astra earlier this month, according to Neurohive.
The livestream and the IPO will test the claims
The commercial pitch is about deployment cost: IT Home reported that Genie Master is meant to let robots improve without the human demonstrations and debugging that new tasks otherwise need. AGIBOT already has scale: China Fund News reported in July that it had started the process for a Hong Kong listing, and chairman Deng Taihua said revenue grew from about RMB300,000 in 2023 to more than RMB1 billion in 2025.
AGIBOT’s reported new funding round, EastFrontier noted earlier this month, could reset its IPO timetable. Policy RSI, which feeds data from working robots back into model training, would also give AGIBOT a growing pool of real-world data.
The November 2 livestream is the first public check. PEdaily reported that the 24 hours will cover the lunchtime and evening peaks, an overnight shift with one robot on duty and the early breakfast rush. The measure that matters is how many tasks the robots finish without human help, and whether AGIBOT publishes before-and-after success rates from the store trial.