DOPPEL: one video world model trained across 8 embodiments, from robot arms to humanoids
_krishna_murthy · x · 2026-11-20
Researchers introduce DOPPEL, a video world model that renders the robot along its planned motion from the camera's viewpoint and conditions a shared video model on those renders plus actions — instead of raw action vectors alone.
- Doppel-8 trains on 600 hours across 8 datasets: robot arms, a bimanual robot, a humanoid, and human hands/bodies
- Beats action-only baselines by 1.00 dB PSNR on unseen trajectories and 1.20 dB out of distribution across three robot-arm datasets; moving-object PSNR rises from 21.50 to 24.85 dB on a held-out toy kitchen
- More sample-efficient: exceeds the shared action-only baseline with 63.8% fewer training draws
- Missing camera poses are recovered from videos; paper and code are available
Related event: Doppel: Open-Source Video World Model for 8 Robot Embodiments(3 posts)→
More from Embodied
- GPT-6-Astra Scores Far Below Published Results When Robot Benchmarks Aren't Tailored — wzenus · 2026-11-20
- Developer claims he 'solved JEPA': model 30x smaller, 100x faster, trains in 30min on a 3080 — MarcJSchmidt · 2026-11-20
- MongoDB founder builds open-source physical AI platform Viam, eyes 1B+ deployments in a decade — gdequeiroz · 2026-11-20
- Waymo robotaxi spotted driving in downtown Toronto at night — GregCook2011 · 2026-11-20
- RealSense Dual RGB + NVIDIA Fast-FoundationStereo live depth demo hits ~8.5 fps — chrismatthieu · 2026-11-20
- Maker builds a voice-command robot servant that figures tasks out itself — Infinite-Job-5616 · 2026-11-20