AgiBot GO-2 (Genie Operator-2)

Architecture

Embodied foundation model with multimodal LLMs + mixture-of-experts

Key facts

GO-2 ViLLA at Partner Conference (primary AGIBOT)

AGIBOT's April 17 2026 Partner Conference release (Shanghai) lists GO-2 (ViLLA Embodied Foundation Model) under Manipulation Intelligence, saying it bridges planning and execution with Action Chain-of-Thought for long-horizon task performance and claiming state-of-the-art results on major benchmarks. Treat SOTA wording as company-stated ([AGIBOT](https://www.agibot.com/article/231/detail/63.html))

Three Intelligences stack (primary AGIBOT)

The same AGIBOT release organizes eight foundational AI products under a One Robotic Body, Three Intelligences architecture (Locomotion, Manipulation, Interactive), naming related models including BFM, GCFM, GE-2, Genie Sim 3.0, SOP, and WITA Omni alongside GO-2 and AGIBOT WORLD 2026. Treat product naming as company-stated ([AGIBOT](https://www.agibot.com/article/231/detail/63.html))

Action CoT and dual-system (primary AGIBOT)

AGIBOT's GO-2 page describes Genie Operator-2 as bridging a semantic-actuation gap via Action Chain-of-Thought (reasoning in action space to emit macro-plans) and an Asynchronous Dual-System with a lower-frequency Semantic Planning Module (System 2) and higher-frequency Action Following Module (System 1). Treat architecture claims as company-stated ([AGIBOT](https://www.agibot.com/article/231/detail/56.html))

Company-stated benchmarks (primary AGIBOT)

The same AGIBOT page reports company-stated benchmark figures for GO-2 including 98.5% average success on LIBERO, 86.6% zero-shot on LIBERO-Plus, 47.4 average on VLABench, and 82.9% real-world success after simulation-only training on Genie Sim 3.0, framed as outperforming pi0.5 and NVIDIA GR00T. Treat scores as company-stated, not independently verified here ([AGIBOT](https://www.agibot.com/article/231/detail/56.html))

Embodied AI foundation model

AgiBot describes Genie Operator-2, or GO-2, as an embodied AI foundation model that connects logical reasoning with physical execution through action intent sequences ([AgiBot GO-2](https://www.agibot.com/article/231/detail/56.html)).

Simulation and benchmarking

AgiBot identifies Genie Sim 3.0 as a simulation, data, and benchmarking platform supporting the evaluation and development of embodied AI systems ([AgiBot Genie Sim 3.0](https://www.agibot.com/article/231/detail/55.html)).

Genie Studio integration

AgiBot's Genie Studio page says the platform integrates base models, data collection, simulation, training, evaluation, and one-click real-machine deployment for embodied-intelligence development, including support for AgiBot's GO-1 foundation model and related model workflows ([Genie Studio](https://genie.agibot.com/en/geniestudio)).

Simulation calibration

AgiBot says Genie Studio uses 3D reconstruction and calibrated simulation environments, and reports less than five percent error between GO-1 model tests and real-device outcomes; treat the metric as company-stated ([Genie Studio](https://genie.agibot.com/en/geniestudio)).

Developed by (1)

• AgiBot

Common questions

Sources (4)

1. https://www.agibot.com/article/231/detail/63.html ↗
2. https://www.agibot.com/article/231/detail/56.html ↗
3. https://www.agibot.com/article/231/detail/55.html ↗
4. https://genie.agibot.com/en/geniestudio ↗

Related brains

• Genie Operator-1 (GO-1)5 models
• Aurora Driver4 models
• Tesla FSD-Bot4 models
• Kodiak Driver3 models
• Nuro Driver3 models
• Redwood AI3 models