Horizon Robotics EmbodiedGen V2 Generative Simulation Infrastructure Launch

Beijing Horizon Information Technology Co Ltd, also known as Horizon Robotics, announced the launch of Horizon Robotics EmbodiedGen V2, a generative simulation infrastructure designed for embodied policy learning. The platform introduces a unified simulation-ready representation that connects task-driven world generation, large-scale scene creation, 3D scene editing, cross-simulator deployment, and closed-loop policy learning within a single framework. The technology aims to improve the development and validation process for intelligent robotic systems by reducing dependence on exclusively real-world training environments.

For asset generation, EmbodiedGen V2 supports multiple pluggable backends, including TRELLIS, SAM3D, and Hunyuan3D. The system accepts text prompts, images, and partially occluded images as inputs while automatically performing quality inspection, mesh repair, simplification, texture baking, physical property assignment, and asset export in URDF, MJCF, and USD formats. Evaluation of 200 generated assets demonstrated a 96.5% human acceptance rate and a 98.6% collision test pass rate, highlighting the platform’s capability for producing simulation-ready digital assets.

EmbodiedGen V2 Advances Robotics Simulation Capabilities

The platform automatically annotates object affordances through functional part segmentation, geometry-consistent post-processing, vision-language-model-guided part merging, and definition of three-dimensional contact regions and actions. Users can modify generated 3D scenes through natural-language instructions, allowing faster iteration during simulation development. The generated content can also be deployed across different simulators by using URDF as an intermediate representation, enabling greater flexibility for robotics researchers and developers working on advanced embodied intelligence applications.

In downstream validation, the research team pretrained the π0 vision-language-action model using real-world data and then performed reinforcement learning exclusively inside generated environments. The approach significantly improved simulation performance, increasing task success rates from 9.7% to 79.8%. Across 12 real-world scenarios and tasks, the model success rate increased from 21.7% to 75.0%, demonstrating the potential of synthetic simulation environments for improving real-world robotic intelligence.

Impact of EmbodiedGen V2 on Future AI Development

The launch of EmbodiedGen V2 reflects growing investment in simulation-driven artificial intelligence development in China and the broader robotics ecosystem. By combining generative AI, 3D asset creation, simulation environments, and reinforcement learning, the platform provides a scalable approach for training intelligent machines. Technologies like this could support future applications across robotics, autonomous systems, and AI-driven mobility solutions where accurate virtual environments are essential for testing complex behaviors before physical deployment.

Frequently Asked Questions

What is Horizon Robotics EmbodiedGen V2?
Horizon Robotics EmbodiedGen V2 is a generative simulation infrastructure designed to support embodied policy learning through advanced virtual environments and AI training methods. The platform integrates world generation, 3D asset creation, scene editing, simulator deployment, and reinforcement learning workflows. It helps researchers develop intelligent systems by creating realistic simulation-ready environments, reducing reliance on physical testing, and improving the ability of AI models to learn complex tasks before deployment in real-world scenarios.