Embodied Video Reward Model
A reward model proposed by DreamTrue for scoring the physical plausibility of embodied video. Its annotation taxonomy covers three categories: robot defects, object defects and interaction defects. The scores it produces are used for reinforcement learning post-training of the video world model.
Timeline (last 90 days)
DreamTrue builds an embodied video reward model to score the physical plausibility of predicted videos, and the scores are used for reinforcement learning post-training.
DreamTrue's reward model training dataset covers three categories of problems—robot defects (such as joint anomalies, self-intersection), object defects (such as objects vanishing into thin air or deforming incorrectly) and interaction defects (such as misaligned hand-object contact points, penetration)—and has undergone manual annotation.
All Facts (2)
DreamTrue builds an embodied video reward model to score the physical plausibility of predicted videos, and the scores are used for reinforcement learning post-training.
50%UnverifiedDreamTrue's reward model training dataset covers three categories of problems—robot defects (such as joint anomalies, self-intersection), object defects (such as objects vanishing into thin air or deforming incorrectly) and interaction defects (such as misaligned hand-object contact points, penetration)—and has undergone manual annotation.
50%