NVIDIA's LidarVerse simulates any LiDAR sensor in real time with one world model
NVIDIAAI · x · 2026-11-17
NVIDIA Research released LidarVerse, the first unified conditional world model for cross-task, cross-sensor LiDAR simulation, with paper and code open-sourced.
- Capabilities: recreates scenes from viewpoints the vehicle never captured, generates scenes from HD maps, and forecasts what a sensor sees next — all in real time
- Architecture: an attention-based range-image tokenizer, a flow-matching DiT teacher initialized from Cosmos-Transfer2.5-2B, and a four-step causal student
- Performance: the distilled causal checkpoint runs at 10 FPS on a single H100; the tokenizer supports 32-, 64-, and 128-beam sensors
- Memory: a rolling nine-frame K/V cache plus a persistent spatial cache enables generalizable resimulation without per-scene optimization
More from Embodied
- Jetson GPIOs now drive $2 L298N motor controllers directly, no Arduino needed — chrismatthieu · 2026-11-17
- UW-led team unveils Grasping the Future, forecasting 3D hand-object motion from internet videos — mangahomanga · 2026-11-17
- New Unitree Robotics datasets land on Hugging Face — mishig25 · 2026-11-17
- Enactra debuts embodied-AI simulation platform where every frame ships annotated — Lianhuiq · 2026-11-17
- LG targets Q1 2027 to unveil bipedal humanoid on NVIDIA reference design — xmercury_one · 2026-11-17
- Robotics researcher: edge cases are literally the entire problem for robot deployment — chris_j_paxton · 2026-11-17