ReaDy-Go: Real-to-Sim Dynamic 3D Gaussian Splatting Simulation for Environment-Specific Visual Navigation with Moving Obstacles

ReaDy-Go introduces a real-to-sim simulation pipeline using 3D Gaussian Splatting to generate photorealistic dynamic environments with moving obstacles for training robust visual navigation policies. The system synthesizes realistic human avatars and motions within reconstructed scenes, enabling policies to better transfer from simulation to real-world deployment across various environments.

ReaDy-Go addresses a critical bottleneck in embodied AI development: the simulation-to-reality transfer gap for visual navigation in dynamic environments. The research tackles two fundamental problems that have limited prior work—the inability to realistically simulate human obstacles and the difficulty of training policies specific to target deployment environments like homes, restaurants, and factories. By combining 3D Gaussian Splatting reconstruction with animatable human avatars and motion synthesis, the approach generates scalable, photorealistic training data without expensive real-world data collection.

This work represents a meaningful evolution in embodied AI simulation methodology. Previous approaches relied either on synthetic assets or static scene representations, limiting the fidelity and diversity of training scenarios. The integration of dynamic human obstacles—derived from 2D trajectory planning—provides substantially more realistic navigation challenges than prior methods. The research demonstrates strong empirical validation through both simulated and real-world experiments, with notable zero-shot generalization to unseen environments.

The broader impact extends beyond academic robotics. Companies developing autonomous systems for indoor navigation, delivery robots, or service robots can leverage this pipeline to dramatically reduce development costs and accelerate deployment timelines. The demonstrated robustness against sim-to-real gaps suggests practical commercial viability. For the AI industry, this represents incremental progress toward more generalizable embodied AI rather than revolutionary breakthrough, as core GS reconstruction and human animation technologies existed prior.

The open-sourced project page indicates potential community adoption. Key observations for tracking include whether this methodology becomes standard in robotics research and whether commercial robotics companies adopt similar GS-based simulation approaches for policy development.

• →ReaDy-Go combines 3D Gaussian Splatting with animatable human avatars to generate photorealistic dynamic navigation scenarios for robust policy training.
• →The pipeline successfully bridges the simulation-to-reality gap, demonstrated through real-world experiments and zero-shot deployment in unseen environments.
• →Dynamic human obstacles synthesized from 2D trajectories provide substantially more realistic training data than static scenes or procedural assets.
• →The approach enables environment-specific policy training for diverse deployment contexts including households, restaurants, and factories.
• →Photorealistic scenario generation from arbitrary viewpoints reduces reliance on expensive real-world data collection for embodied AI development.