The fields of 3D reconstruction and generative video are moving at a staggering pace. Over 1,700 3D Gaussian splatting papers have been published since mid-2023, and video diffusion research has surpassed 2,300 papers in roughly the same window. The state of the art is changing month to month. Techniques that were considered standard, like COLMAP for structure from motion, are hitting fundamental limits in speed and robustness that new deep-learning-based approaches are rapidly overcoming. Parallel Domain has a dedicated team working at the frontier of these methods, and in this talk we share what we are seeing ahead: how 3D Gaussian splatting, feed-forward reconstruction models and proprietary pose estimation are converging to turn a single drive log into a fully controllable, photorealistic 4D simulation, bridging the gap between rigid replay and hallucinated generative video to enable software-augmented testing at scale.

• How research in 3D Gaussian splatting and video diffusion has exploded, and why the state of the art in 3D reconstruction is shifting faster than ever
• Why industry-standard tools like COLMAP struggle with speed and sparse-feature environments, and how new deep-learning approaches to structure from motion are overcoming these limits
• How a single real-world drive log can be transformed into a controllable, sensor-realistic 4D simulation
• What this shift means for engineering and program leaders evaluating simulation strategies: how software-augmented testing reduces reliance on costly real-world fleet miles while accelerating safety validation timelines