LidarVerse: one real-time LiDAR world model for reconstruction, forecasting and generation
ZGojcic · x · 2026-10-20
Researchers including Sanja Fidler and Laura Leal-Taixé present LidarVerse (arXiv:2610.19201), a unified real-time LiDAR world model across tasks and sensors. A universal tokenizer maps range sequences into a shared latent space where a bidirectional teacher conditions on LiDAR prefixes, HD-maps, a persistent spatial cache and optional text; varying available evidence switches the same model between world generation, point-cloud forecasting, and generalizable re-simulation without per-scene optimization. Distilled into a 4-step autoregressive student with rolling temporal K/V and world-aligned spatial caches, it runs at 11 FPS on a single H100 and achieves SOTA on most metrics, supporting 32/64/128-beam LiDAR. Paper, code and models are open.
More from Embodied
- Surgeons Complete 100+ Cataract Surgeries Using Apple Vision Pro and ClearSphere App — asymco · 2026-10-20
- What makes human video generation good for robot manipulation? A constrained-optimization analysis — mangahomanga · 2026-10-20
- Figure's F.03 humanoid walks 348 miles and 1.2M steps over 9 straight days on a treadmill — adcock_brett · 2026-10-20
- NVIDIA joins Gemma hardware hackathon with 30 Jetson kits at AGI House on Oct 24 — agihouse_org · 2026-10-20
- Tesla shows Optimus V3 render in investor letter, Musk touts an 'Amazing Abundance' era — JOBhakdi · 2026-10-20
- UCSD PhD Nicklas Hansen joins NYU as postdoc with LeCun and Pinto on world models + robots — LerrelPinto · 2026-10-20