Summary

Genie Sim 3.0 provides a comprehensive simulation platform for embodied intelligence development, addressing environment reconstruction, scene generalization, data collection, and automated evaluation. Targeting researchers and developers, it accelerates model development and reduces hardware reliance through high-fidelity assets, LLM-driven scene generation, and extensive synthetic datasets.

How It Works

The platform integrates a 3DGS-based reconstruction pipeline with visual generative models for high-fidelity environments. It pioneers LLM-driven scene generation, enabling rapid creation of diverse scenarios and evaluation configurations via natural language. A VLM-based auto-evaluation system profiles model capabilities, complemented by over 10,000 hours of synthetic data for loco-manipulation tasks. This approach aims for efficient, scalable, and accurate embodied AI development.

Quick Start & Requirements

Installation, asset, and dataset access are detailed via provided links to a User Guide, Assets, and Dataset pages. Updates mention support for Isaac Sim v5.1.0 and RTX 50 series GPUs, suggesting a requirement for capable hardware.

Highlighted Details

• High-Fidelity Assets: 5,140 validated 3D assets across five real-world operation fields.
• LLM-Driven Generation: Instant scene and evaluation configuration generation via conversational interaction.
• Large-Scale Synthetic Data: Over 10,000 hours of open-source data for 200+ loco-manipulation tasks.
• Benchmark & Evaluation: 100,000+ scenarios with LLM-generated instructions; claims <10% real-world discrepancy.
• Sim-to-Real Transfer: Models trained with Genie Sim data show superior zero-shot sim-to-real capabilities.

Maintenance & Community

Active development is indicated by recent updates, including v3.0 (Jan 2026). The roadmap lists pending features like Docker support for reconstruction and Huggingface uploads. No direct community links are provided.

Licensing & Compatibility

Code and data in source/geniesim and source/data_collection are under Mozilla Public License 2.0 (MPL 2.0). The source/scene_reconstruction project has mixed licenses requiring review of individual LICENSE files. MPL 2.0 is generally permissive for commercial use but mandates sharing modifications.

Limitations & Caveats

Pending roadmap items include Docker support for the scene reconstruction pipeline and Huggingface uploads. The mixed licensing in the scene_reconstruction component requires careful compatibility assessment.