Many users are excited about Xiaomi's 38B world model for embodied robotics because synthetic data enables major progress that real-world demos cannot match, while others worry it risks model collapse and unpredictable robot policies.
Based on 8 visible X reactions from 22 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@Scobleizer @kylecompute Also different domain here. Synthetic data in the LLM world causes model collapse. ENV data or physical based words has proven more reliable. Even then Teslas virtual ENV are heavily derived through gaussian splats to pin them to “reality
@kylecompute @Scobleizer @kylecompute Robot intelligence = LOT of data 🤖 You can't collect enough real-world demos manually. Synthetic generation with multi-view consistency is the path. That OOD jump to 63.2% is massive.
@Scobleizer The scary part is not the 38B model, it is synthetic data finally transferring this well into real robot policies.
@teortaxesTex Bytdanace cooking. They are also preparing a large LLM. Probably seedance 2.0 moment for text and coding.
It raises real-world robot task success from 36.9% to 63.2%.
@Scobleizer @kylecompute Also different domain here. Synthetic data in the LLM world causes model collapse. ENV data or physical based words has proven more reliable. Even then Teslas virtual ENV are heavily derived through gaussian splats to pin them to “reality
@kylecompute @Scobleizer @kylecompute Robot intelligence = LOT of data 🤖 You can't collect enough real-world demos manually. Synthetic generation with multi-view consistency is the path. That OOD jump to 63.2% is massive.
@Scobleizer The scary part is not the 38B model, it is synthetic data finally transferring this well into real robot policies.
@teortaxesTex Bytdanace cooking. They are also preparing a large LLM. Probably seedance 2.0 moment for text and coding.
@teortaxesTex 这玩意儿确实猛啊
@Scobleizer Thanks Robert
And if you wanted to ask: yes, it is" Anyway, interesting that we have 2 near-identical models derived from Emu 3.5 (this and ByteDance Seed's UniVR) in such rapid succession. https://x.com/teortaxesTex/status/2077059955320906110/photo/1 https://twitter.com/sheriyuo/status/2076920174205345831
Robot breakthrough. Just what I didn’t want to see from China. “Its synthetic data significantly boosts real-robot policy performance, lifting out-of-distribution task success from 36.9% to 63.2%” https://twitter.com/sheriyuo/status/2076920174205345831
Many users are excited about Xiaomi's 38B world model for embodied robotics because synthetic data enables major progress that real-world demos cannot match, while others worry it risks model collapse and unpredictable robot policies.
Based on 8 visible X reactions from 22 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@Scobleizer Thanks Robert
And if you wanted to ask: yes, it is" Anyway, interesting that we have 2 near-identical models derived from Emu 3.5 (this and ByteDance Seed's UniVR) in such rapid succession. https://x.com/teortaxesTex/status/2077059955320906110/photo/1 https://twitter.com/sheriyuo/status/2076920174205345831