DyRAD: radar novel view synthesis for dynamic driving scenes, recovering 90.7% vs 26.9% baseline
Technion · hf · 2026-10-01
Technion researchers present DyRAD, the first method to render full range-azimuth-Doppler (RAD) tensors for dynamic driving scenes.
- Prior radar novel-view synthesis either renders Doppler only for static scenes or reconstructs just range-azimuth tensors for dynamic ones.
- DyRAD models scenes with static background reflectors plus motion-tracked dynamic point reflectors; Doppler serves as both rendered output and supervision for object tracks.
- Reflectors are rendered through a fixed analytic point-spread function derived from the radar's signal-processing chain, preventing sensor-induced spread from contaminating scene geometry — also enabling zero-shot transfer across radar configurations.
- On RADIal, DyRAD recovers radar detections in 90.7% of reference-detected objects vs 26.9% for the strongest baseline; validated on Boreas and a synthetic benchmark, including displaced viewpoints untested by prior work.
More from Embodied
- Personalized AI, personalized bugs: Dots users report wildly different flaws — ChrisGPT · 2026-10-01
- A 40-gram autonomous drone hunts mosquitoes by acoustic signature and sonar — TinfoilTricorn · 2026-10-01
- Google DeepMind Tokyo hiring senior research engineer for gaming AI agents — heiga_zen · 2026-10-01
- 5 robotics accounts that are '6 months ahead': Jim Fan, Physical Intelligence founders lead the list — RemiCadene · 2026-10-01
- VLA reinforcement learning updates are low-rank and concentrate in overlooked Timestep Modules — HOLILAB · 2026-10-01
- NVIDIA's Ming-Yu Liu on open models, world models, and the future of physical AI — Practical AI · 2026-10-01