Odyssey 发布世界模型并展示机器人与驾驶策略应用

AI 导读

Odyssey 将世界模型中的物理预测与语言模型中的编码相类比，认为能准确预测物体下落、倾倒和碰撞的模型体现了机器人和智能体所需的世界理解。文中提到机器人手臂从数十小时演示中学习任务，使用冻结的 Odyssey-3 骨干在 20 小时驾驶数据上训练的驾驶策略，以及 Flexion 的人形策略在让 VLA 基线失效的光照变化下仍能工作。

正文 · 原文

Odyssey draws a parallel between physics in world models and coding in language models. A model that predicts falling, pouring, and colliding objects accurately shows an understanding of the world, which robots and agents require.

Robot arms learned tasks from tens of hours of demonstrations. The driving policy trained on 20 hours of driving data with the Odyssey-3 backbone kept frozen. Flexion's humanoid policies continued to work under lighting changes that broke the VLA baselines Odyssey tested.

Physical AI teams can contact Odyssey to build on the model.

Test it out 👀

experience.odyssey.systems/