AGIBOT GO-3 Robot Model: Learning on the Job in a Store

Chinese robotics company AGIBOT has introduced GO-3 Preview, an AI model meant to help robots keep improving after they leave the lab. According to Interesting Engineering, which covered the launch, the headline result is a robot in a real convenience store raising its success rate from 28.6 percent on day one to a stable 85.7 percent after five days. GO-3 Preview launched on October 29. That is a company claim, and the details matter more than the percentage.

What AGIBOT announced

GO-3 Preview is built on what AGIBOT calls a World Reasoning Action Model, or WRAM. As described in the article, it links three things: reasoning about a task, predicting what will happen, and producing the movements to carry it out.

AGIBOT says it trains the model on two kinds of data. One is first-person video of people handling objects and doing everyday activities. The other is robot trajectories, which tie what a robot sees to the actions it took and what resulted. The company says the aim is for robots to apply what they have learned to tasks they have not seen before.

AGIBOT also says the model has taken first place on three embodied AI benchmarks: RoboDojo, RoboTwin and RoboColiseum. The article does not explain what those tests measure, so treat the rankings as the company's report for now. AGIBOT plans to release the full model as open source by the end of 2026.

The convenience store test

The most relevant part for anyone thinking about robots at work or at home is the deployment. AGIBOT says it put a G2 robot in a real convenience store and used its Genie Master agent framework, which connects decision-making, task execution and feedback. According to the company, the robot worked out why it failed, changed its approach and stored what it learned, without human demonstrations or manual tuning.

The reported result is a jump from 28.6 percent to 85.7 percent success across the complete task workflow over five days.

Several things are not stated in the article:

• What the task workflow actually was.
• How many attempts those percentages are based on.
• Whether anyone supervised or reset the robot between attempts.
• What happened to the remaining 14.3 percent of attempts, and whether a failure in a store has any safety or cost consequences.

An 85.7 percent success rate is also well short of what most people would accept from a machine handling their dishes or laundry unattended.

Other demos and the training system

For post-training, AGIBOT built a system called Genie Evolver. The company says it combines data from roughly 100 physical robots with physics simulation and world models, so robots can practice, assess their results and try again.

AGIBOT offered a few smaller examples:

• In tabletop curling, a robot beat the average of the humans tested after four hours of online practice.
• In a block-tower game, a robot pulled out five blocks in a row without the tower falling.
• On a task involving hanging small hooks, AGIBOT reported a 17.7-percentage-point gain in single-task success over conventional supervised fine-tuning.
• The model can learn multi-step tasks such as making jianbing, a Chinese savory crepe, from a small number of demonstrations, according to the company.
• After fine-tuning, it can assemble Lego pieces, with an AI agent coordinating the steps.

These are company-selected demonstrations. The article does not say how they were run, how often they succeeded, or whether the home and kitchen settings it mentions were real households or test setups.

What it means for your home

The idea here is useful: a robot that learns from its own mistakes on the job could, in principle, need less custom programming for each new house and each new chore. Homes are messy and varied, and that is exactly where fixed, pre-trained skills tend to break down.

But this announcement is a research preview, not a product. The article gives no price, no ship date, and no word on whether any robot using GO-3 will be sold for household use. The one real-world trial was in a convenience store, and its task is not described. Nothing in the report speaks to privacy or safety in a living space, such as what a robot that keeps learning stores about your home.

The planned open-source release could be the more practical development. If it happens as AGIBOT describes, outside researchers and developers could test the claims themselves, which would tell us more than a single company's numbers. Until then, the sensible reading is that learning on the job is a promising direction, and that a home robot you can trust with everyday chores is still not something this announcement puts on a calendar.