Verne V0.5 trains a robot policy with video prediction, but runs actions-only
rohanpaul_ai · x · 2026-11-06
Verne V0.5 is a "World Action Model" (WAM) trained on two objectives at once: predicting the robot's next actions and predicting what its cameras will see next.
- Video prediction exists only to teach the model how scenes change; during training, a video transformer and an action transformer exchange information via masked attention, each with its own flow-matching objective (related work: Fast-WAM, Motus, VPP2)
- At runtime, the model outputs actions directly, with no per-step video generation
Design insight: distill a world model's scene understanding into the policy without paying the inference cost of generating video at every control step — a representative recipe for combining world models with robot policies.
More from Embodied
- Google Launches Fitbit Edge: 50m Water Resistance and First User-Replaceable Battery — GoogleDesign · 2026-11-06
- Meta lets US users video chat with their Muse AI avatar ahead of wearable launch — TechCrunch AI · 2026-11-06
- Next year will be the year of robotics evaluation — cool demos alone won't raise funding — DJiafei · 2026-11-06
- eigenrobot poses with Tesla Optimus humanoid at live event — beffjezos · 2026-11-06
- Giving Claude a cutting-edge robot body just so it can spoon-feed me — repligate · 2026-11-06
- GPT-6 Astra Tops Robot Leaderboard but Went 0/11 on Real Plug Insertions — AIFlow_ML · 2026-11-06