Lu Zongqing, a professor at Peking University and founder of BeingBeyond, is pioneering a "latent space" approach to embodied AI. This method bypasses visual generation, directly predicting robot actions and world responses in an embedding space, significantly reducing training costs and enabling real-time control. Lu believes that a deterministic technical paradigm, similar to what enabled large language models, is still missing for embodied AI, and that first-person human videos are the most scalable data source for training these models. BeingBeyond recently released Being-H0.8, an implicit haptic world-action model that integrates visual, haptic, and action modalities into a unified latent space. AI
IMPACT This latent space approach could significantly reduce the cost and increase the speed of embodied AI development, potentially accelerating real-world robot deployment.
RANK_REASON The article discusses a new technical approach and model release in embodied AI research by a university professor and his startup. [lever_c_demoted from research: ic=1 ai=1.0]
- BeingBeyond
- Being-H0.8
- NeurIPS
- Cosmos Policy
- ICLR
- Latent World-Action Model
- Li Feifei
- Lu Zongqing
- Next Token Prediction
- NVIDIA
- Peking University
- Transformer
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
