🧠 Food for thought

Implications, context, and why it matters.

Humanoid AI is a full-stack challenge, not a single model

• Software that runs a humanoid robot needs a layered stack, not one large model.
• NVIDIA’s Isaac GR00T N1.6 runs a multimodal Vision-Language-Action (VLA) model for high-level instructions. It uses world models such as NVIDIA Cosmos Reason to turn those instructions into step-by-step plans based on the surrounding scene, plus separate modules for whole-body control, navigation, and localization 1.
• Delta Intelligence pairs foundation models with 3D world modeling software, which lines up with this full-stack direction. The broader research area includes dedicated surveys and benchmark ecosystems 2.

Delta’s funding adds to China’s broader push around “embodied intelligence”

• Delta Intelligence’s funding sits within China’s wider push for “embodied intelligence,” which refers to AI systems embedded in physical machines such as robots 3.
• An industry group for this work already operates in the Yangtze River Delta. The region produces one-third of China’s AI industrial output 3.
• State support includes Shanghai grants of up to 50 million yuan (US$7.33 million) for certain industrial-cluster projects tied to the G60 science and technology innovation corridor measures 4.
• China is incubating startups like Delta Intelligence and drawing investment from robotics original equipment manufacturers (OEMs). This helps grow a domestic software ecosystem that competes with integrated platforms such as NVIDIA’s Isaac GR00T and Cosmos 5.