NVIDIA's Foundation4D unifies driving perception and 4D reconstruction in one feed-forward model
zhenjun_zhao · x · 2026-10-24
NVIDIA released Foundation4D, a feed-forward model that unifies multi-camera driving perception and 4D reconstruction.
- Shared DINOv2 encoder with factorized image/view/time attention over 6-camera rigs
- Task-specific heads output rig/ego pose, metric depth and semantic point maps, 3D Gaussians, cuboids with velocities, and BEV semantic occupancy
- Sets strong depth results on Waymo/nuScenes (AbsRel 0.09/0.15, δ<1.25 of 91.97%/86.44%), well ahead of DVGT and GeUP baselines, plus large trajectory-error gains
- Supports a PGO+LM inference mode reconstructing 20s six-camera clips into RGB and semantic point clouds
More from Embodied
- NVIDIA expands Isaac robotics platform at GTC Berlin; enact_intel stealths with $30M pre-seed — lukas_m_ziegler · 2026-10-24
- Robotiq launches 2FC adaptive gripper with embedded tactile intelligence, ships Q1 2027 — lukas_m_ziegler · 2026-10-24
- Reliable Robotics raises $115M at unicorn valuation for autonomous Cessna flights — lukas_m_ziegler · 2026-10-24
- Eka Robotics raises $150M at $1.45B valuation; vision-force-action picks 1,500 objects/hour — lukas_m_ziegler · 2026-10-24
- Robotics weekly roundup: NVIDIA, Agile Robots, Eka Robotics and more — lukas_m_ziegler · 2026-10-24
- KAIST's FRED Self-Distills Single-Step Video Representations, Boosting Robot WAMs on Real Tasks — coallaoh · 2026-10-24