AgiBot releases GO-1

Background

Most robots in factories repeat one motion at one station, programmed once and rarely changed. A robot meant to work around people has to handle tasks, objects and rooms nobody wrote a program for, which is why many teams have moved to learning-based methods: instead of encoding rules by hand, they train a model on data and let the robot adapt.

GO-1 comes out of that setting. It was released by AgiBot, which China News Service described as Shanghai's first mass-production humanoid robot company^[1]. AgiBot said that by January 2025 it had accumulated 1,000 mass-produced general-purpose embodied robots^[2]. "Embodied" here means intelligence that acts through a physical body rather than only producing text or images.

The release

On 10 March 2025 AgiBot released GO-1, presented by the company as a general-purpose embodied foundation model^[1][2]. A foundation model in this sense is trained to serve as a common base: many downstream skills are built on top of it, instead of each one getting its own separately trained system.

What the company says it can do

AgiBot said GO-1 can be deployed across several of its robot bodies and can keep learning from data collected during actual use^[2]. The second point addresses a familiar gap between demonstration and deployment: a policy tuned in a lab meets slightly different lighting, objects and surfaces in a real setting, and a model that keeps updating from the data generated while working does not have to be sent back for retraining every time.

The company also said the model supports learning from human videos, few-shot generalisation to new tasks, and transfer across robot forms^[2]. Learning from human video matters because footage of people doing chores or moving objects is cheaper to collect than robot demonstrations. Few-shot generalisation means a new task should need a handful of examples rather than thousands of trials. Transfer across forms means something learned on one body can be used on another — the condition that makes a shared "base" worth having for a company that builds more than one kind of machine.

The longer timeline

Model progress is not the same as robots arriving in homes. Ren Guanghui, a researcher at AgiBot, said the technology was at a transition stage from G3 to G4, and that humanoid robots entering homes and creating value there would take about five more years^[1]. Set against that, GO-1 is described as already deployable across several of the company's robot bodies^[2]. Between running on a company's own machines and doing useful work in a household there is still a stretch of time, and the two horizons in the announcement — a model available now, homes about five years out — are the most concrete part of the story.