AGIBOT GO-3 Preview

What GO-3 Preview is

GO-3 Preview is an embodied AI model from AGIBOT: a model meant to let a robot understand a task, reason about the physical scene in front of it and produce the actions that carry the task out. The company showed it at its Innovation Day in Shanghai on 29 October 2026.^[1] The name marks it as a preview — a version published ahead of a final release.

Embodied models differ from the ones behind chat assistants in an obvious way. Their output is not text but movement: a sequence of commands that makes a machine reach, grasp, walk or assemble. That makes the problem harder in kind, because the model has to keep track of a world that changes while it is acting, and because a wrong output has physical consequences rather than just a bad sentence.

A world reasoning action model at the core

AGIBOT says the model is built on WRAM, a World Reasoning Action Model that unifies task understanding, prediction of future states and action generation.^[1] In many robot systems those three jobs sit in separate modules: one interprets the instruction, another models how the scene will evolve, a third turns the plan into motor commands. WRAM's premise is that they can be handled inside a single model, so that what the robot expects to happen and what it decides to do are produced together rather than stitched together afterwards.

AGIBOT names three designs — CoAE, SVP and KASO — that connect what must be done, what may happen next and the actions needed to finish a task.^[1]

A learning system around the model

GO-3 Preview is presented as part of an integrated learning system rather than as a standalone network. Two other components sit alongside it: Genie Evolver, a hybrid post-training system, and Genie Master, an embodied agent framework.^[1] Roughly, pre-training gives a model its general competence, post-training shapes it toward particular tasks and standards, and an agent framework governs how it behaves over a long stretch of work rather than a single step.

AGIBOT calls the overall approach full-lifecycle scaling, because it spans pre-training, post-training and deployment in the real world.^[1] In its Innovation Day material the company also presents an approach called Physical Intelligence Scaling, which combines pre-training, post-training and continuous improvement after deployment.^[2] The emphasis in both descriptions falls on the same idea: capability is not finished when training ends, and what a deployed robot encounters is treated as further input to the system.

Benchmark results

AGIBOT says GO-3 Preview took first place on three embodied AI benchmarks: RoboDojo, RoboTwin and RoboColiseum.^[1] The split of results is specific. On RoboDojo the model ranked first in both task success rate and process score, while on RoboTwin and RoboColiseum it recorded the highest overall score.^[2]

Success rate is the straightforward measure — the share of tasks a model completes. Process score looks at the path taken to completion rather than only the endpoint, which matters for robots because two runs can end in the same place while one of them knocked things over on the way. The figures above are those reported by AGIBOT in its own announcement.

Zero-shot and few-shot behaviour

AGIBOT demonstrated zero-shot and few-shot capabilities with the model.^[2] Zero-shot means attempting a task without task-specific examples; few-shot means attempting it after seeing a small number of demonstrations. Both are tests of whether a model has learned something general about how the physical world behaves, as opposed to memorising a fixed repertoire of motions. For robots this is the difference between a machine that can only repeat rehearsed routines and one that can be pointed at a new object or a new arrangement of a workspace.

Open source and where it sits

AGIBOT says GO-3 Preview will be fully open-sourced by the end of 2026.^[1] Full open-sourcing of a model of this kind means other groups can run it, test it on their own tasks and compare results against the published benchmark figures.

The model belongs to a generation of systems that treat a robot's brain as a large learned model rather than a hand-written pipeline of perception, planning and control. Its distinguishing pieces, as described by the company, are the fusion of prediction and action inside WRAM, and the decision to build training and deployment into one loop. What the preview naming and the open-source plan leave open is how much of that survives contact with hardware in ordinary use.