World models are becoming a better and better way to train policies; they're also increasingly a solid way to assess the quality of policies.
🚀 Introducing 𝚆𝙴𝙰𝚅𝙴𝚁 🌎, a world model for robot manipulation built around three key desiderata: (i) 🔍 Fidelity, (ii) 🔗 Consistency, and (iii) ⚡ Efficiency. 𝚆𝙴𝙰𝚅𝙴𝚁 unlocks multiple downstream applications—policy evaluation, policy improvement, and test-time planning. 🧵
No Digg Deeper questions have been answered for this story yet.
@chris_j_paxton Yeah, but they still require high-end hardware, and even this or Flash-WAM below is actually kind of far from real-time inference performance. Policy training is cool, but I thought WAMs would be the king of test-time compute
We need more world models built with real robot use-cases in mind! Check out Arnav and Yilin's fast and consistent world model for manipulation -- the model weights, code, and paper are all open-sourced!
🚀 Introducing 𝚆𝙴𝙰𝚅𝙴𝚁 🌎, a world model for robot manipulation built around three key desiderata: (i) 🔍 Fidelity, (ii) 🔗 Consistency, and (iii) ⚡ Efficiency. 𝚆𝙴𝙰𝚅𝙴𝚁 unlocks multiple downstream applications—policy evaluation, policy improvement, and test-time planning. 🧵
