NVIDIA AI Updates: April 17, 2026

1. Lyra 2.0 Open-Sources Walkable 3D Worlds from a Single Photo

NVIDIA. NVIDIA’s Spatial Intelligence Lab released Lyra 2.0, a system that generates large, coherent 3D environments from a single photograph and exports them as Gaussian splats and meshes ready for Isaac Sim or game engines. Two technical contributions stand out: spatial memory that keeps 3D geometry consistent across camera reversals (fixing “spatial forgetting”) and self-enhanced training that corrects cumulative drift over long trajectories. Weights and code are on Hugging Face and GitHub under Apache 2.0, enabling commercial robotics use. The practical payoff is dramatically cheaper synthetic training environments for embodied AI. Source

2. DeepStream 9 Ships Vision AI Pipelines Built by Coding Agents

NVIDIA. A new developer blog details how DeepStream 9 integrates with Claude Code and Cursor to generate real-time vision AI pipelines from natural-language specs. The workflow produces deployable, optimized code for stream ingestion, inference, and analytics, reducing the hand-written plumbing traditionally required to wire GStreamer, TensorRT, and tracker components together. This is a concrete example of coding agents moving from general app development into specialized systems programming. Source