AgiBot GO-2

AgiBot's 2026 second-generation embodied foundation model: it first plans a coarse action sequence, then executes it.

GO-2 is the embodied foundation model AgiBot released on April 17, 2026, at its “2026 Partner Conference” in Shanghai — the successor to 2025's GO-1, which the company describes as a ViLLA-style embodied foundation model. Two ideas are central. First, “action chain-of-thought”: instead of first writing text or generating an image and only then deriving actions, the model directly generates a sequence of coarse-grained action intentions in action space as a high-level plan; the related paper, ACoT-VLA, was accepted at CVPR 2026. Second, an asynchronous two-system design: a low-frequency semantic planning module produces the plan, and a high-frequency action-following module combines it with real-time observations to output control signals — this part was accepted at ACL 2026. AgiBot reports an average LIBERO success rate of 98.5% and 86.6% zero-shot on LIBERO-Plus. Released alongside it was GE-2, a world-action model used to test policies in virtual environments.

ExampleAgiBot reports that GO-2, trained only on simulated data, reached an 82.9% success rate in real-robot testing.

Also called

Genie Operator-2, GO-2 (ViLLA)

Related

AgiBot GO-1 · Vision-Language-Latent-Action · Action Chain-of-Thought · Dual-System Architecture (System 1 / System 2) · AgiBot · LIBERO Benchmark

Sources

AGIBOT Unveils Genie Operator-2 (GO-2) Next-Gen Embodied Foundation Model
AGIBOT Unveils New Generation of Embodied AI Robots and Models (APC 2026, 2026-04-17)
ACoT-VLA (AgibotTech GitHub)

As of

2026-04