Video-Point Model learns 4D world models via 3D point trajectory forecasting
Byungwoo Jeon · hf · 2026-10-21
Video-Point Model (VPM) predicts future multi-view videos and persistent 3D point trajectories from multi-view video histories, camera parameters, and language instructions, improving both video forecasting and robot control through shared 3D-aware representations.
- Ships with VP-WorldBench and VPM-Data spanning real and simulated scenes: tabletop robot manipulation and urban dynamics (humans/vehicles).
- Beats Wan2.2-5B, Cosmos3-Nano, HunyuanVideo-1.5, and X-WAM on FVD; ablations show track supervision (V+T) cuts FVD from 78.52→67.35 (robot) and 101.37→81.92 (general).
More from Embodied
- Menlo Research open-sources Cyclotron, an RL training framework for its Asimov humanoid robot — freelerobot · 2026-10-21
- Cactus runs local speech-to-actions on a $10 ESP32 with sub-30MB models — alexcovo_eth · 2026-10-21
- 3 Robotics Startups Launched in 2 Months — All Just Wrapping the Unitree G1 — saranormous · 2026-10-21
- Berkeley paper: first tractable estimator of empowerment from offline data for exploration — svlevine · 2026-10-21
- Musk says Neuralink will implant first vision devices in 6-12 months, even for the blind from birth — mitchdeg · 2026-10-21
- If humanoid robots make most goods, why should profits belong only to their owners? — VraserX · 2026-10-21