AgiBot GO-1 (Genie Operator-1)

What it is

GO-1, short for Genie Operator-1, is a model that AgiBot released on 10 March 2025 and described as a general-purpose embodied model foundation ^[1][2]. The framing matters: it is presented not as a program written for one specific job but as a base that other robot tasks can be built on top of ^[1][2].

Embodied here means the model does more than handle text and images on a screen; it produces actions that move a physical body. That sets two hard problems at once. The model has to perceive a scene and understand an instruction, and it has to translate "what should happen" into "how the joints and hands should move." GO-1's design is an attempt to split those problems into parts that can be trained on different kinds of data.

The ViLLA architecture

GO-1 is built on an architecture AgiBot calls ViLLA — Vision-Language-Latent-Action ^[1][2]. Between the usual vision-language model and the action output, it inserts a layer of latent actions, and that part is a mixture-of-experts module containing a Latent Planner and an Action Expert ^[1][2]. A mixture-of-experts module is a common way of organising a network so that different sub-parts handle different inputs rather than every input travelling the same path.

The three components divide the work.

The vision-language model gains general scene perception and language understanding from massive internet image-text data ^[1]. The value of this layer is breadth: internet data covers an enormous range of objects, rooms and phrasings, so the model is less likely to be completely lost in a kitchen, a pile of clothes or a container it has never been trained on directly.

The Latent Planner learns general action understanding from large amounts of cross-embodiment and human manipulation data ^[1]. Cross-embodiment means data from robots of different shapes — different arm counts, joint layouts and gripper types. Trained across them, the model is pushed toward the logic of a task rather than the habits of one particular machine. Human manipulation data, such as recordings of people folding or hanging clothes, is far more plentiful than robot data, so including it widens the pool the model can learn from.

The Action Expert learns fine action execution from millions of real-robot data points ^[1]. Real-robot data is the closest to the conditions of deployment: real motors, real friction, real sensor noise. It shows how an action actually lands as a concrete command.

AgiBot says the three components together let the model learn from human videos and generalise quickly from few samples ^[1].

Results

AgiBot tested GO-1 on five tasks of varying complexity and reported that average success rate rose by 32%, from 46% to 78% ^[2]. In one comparison, adding the Latent Planner alone raised success rate by 12%, from 66% to 78% ^[2]. Numbers of this kind matter because they attribute the gain to a specific part of the architecture rather than to the model simply being larger.

Yao Maoqing, an AgiBot partner and president of its embodied intelligence business, said the model can infer the steps needed for a task such as hanging clothes and then execute them in sequence ^[1].

The data underneath

GO-1 was built on AgiBot World, a dataset the company says contains more than 1 million trajectories across 217 tasks and five categories of scenes ^[2]. A trajectory is the recorded sequence of actions a robot produces while completing a task. The number of tasks sets the breadth of coverage; the scene categories set the variety; together they determine how rich a vocabulary of actions the model has seen.

Where it sits

Calling GO-1 a foundation emphasises reuse: one model as a starting point, adapted to particular jobs with relatively little extra data. The word is borrowed from practice in large language models, where a general model is trained on broad data first and then fitted to specific uses. GO-1 carries that idea over to robot action, and assembles its training material from three sources — cross-embodiment and human data for planning, and real-robot data for execution ^[1][2].

The reported success rates come from AgiBot's own tests on five tasks ^[2]. How the model holds up across a wider and messier range of jobs is a separate question, and one that the published figures do not answer on their own.