[Podcast] Robot Learns by Watching
Автор: Vinh Nguyen
Загружено: 2026-01-26
Просмотров: 9
Описание:
Cosmos Policy: Fine-Tuning Video Models for Robotic Control
https://www.alphaxiv.org/abs/2601.16163
Cosmos Policy is a sophisticated robotics framework that adapts the NVIDIA Cosmos video foundation model into a versatile tool for visuomotor control and planning. By employing a technique called latent frame injection, the system integrates robot actions, proprioception, and state values directly into the video model's existing architecture without needing structural modifications. This unified approach allows the model to function simultaneously as a policy, world model, and value function, enabling it to "imagine" future outcomes and select the most successful actions. Experimental results demonstrate that Cosmos Policy achieves state-of-the-art performance in both simulated benchmarks and real-world bimanual manipulation tasks. Furthermore, the model can be refined through policy rollouts, significantly improving its ability to perform high-precision tasks by learning from its own physical experiences.
#ai #nvidia #robot #robotics #standford #research #cosmos #reinforcementlearning
Повторяем попытку...
Доступные форматы для скачивания:
Скачать видео
-
Информация по загрузке: