- Developers can generate interactive 3D worlds in minutes using Genie Sim 3.0 by simply inputting text or images.
- The platform aims to address the data drought in the embodied AI industry and accelerate the commercial rollout of robots.
Chinese humanoid robot startup Agibot launched Genie Sim 3.0, a new simulation platform that allows users to generate interactive and trainable 3D worlds using natural language or images.
The technology shifts environment construction from manual building to instant generation, boosting research and development efficiency for embodied AI, according to an Agibot statement on Wednesday.
By inputting a single sentence or image, developers can prompt the system to build an interactive, navigable 3D world within minutes.
This "input-as-scene" model eliminates the need for manual 3D modeling, leveraging large multimodal models to achieve infinite generalization of simulation environments, the statement said.
The platform also introduced a simulation benchmark covering five core robotic capabilities, supporting the comprehensive evaluation of mainstream foundation models in complex scenarios.
Models trained purely on simulated data can achieve zero-shot transfer to the real physical world, with performance discrepancies of less than 10%, the announcement said.
Furthermore, the platform fully supports the RLinf framework, decoupling its physics and rendering engines to provide high-precision physical simulations at up to 1,000Hz, Agibot said.
This addresses the shortcomings of existing large models and provides a complete toolchain for embodied AI, ranging from efficient training to closed-loop evaluation, the company said.
The upgrade creates synergy with the Agibot World 2026 dataset open-sourced by the company on Tuesday. The platform solves scene construction challenges while the dataset provides real-world data, jointly pushing embodied AI technology from the lab toward commercialization.
