Hydra-0: Action Flow for Generalist World Modeling and Control
Hydra-0 turns robot actions into pixel motion so one model can reason across different robots, tasks, and video backbones.
The paper says its best setup cuts robot-motion error by 90.4% and object-motion error by 60.2% against an action-conditioned baseline. On RoboLab, replayed success rates closely match reference success rates, with Pearson r=0.96. The authors also report an inverse mode that can infer compatible robot motion from desired object flow taken from a human demonstration, then map latent features to executable actions without task-specific expert robot demos. HF Daily Papers' note
The paper says its best setup cuts robot-motion error by 90.4% and object-motion error by 60.2% against an action-conditioned baseline. On RoboLab, replayed success rates closely match reference success rates, with Pearson r=0.96. The authors also report an inverse mode that can infer compatible robot motion from desired object flow taken from a human demonstration, then map latent features to executable actions without task-specific expert robot demos. HF Daily Papers' note
score 6