Fei-Fei Li on Spatial Intelligence: AI Needs Touch and Simulation, Not Just Vision

APPSO · wechat · 2026-08-01

The article explores the evolution of AI from mere image recognition to 'spatial intelligence' and embodied AI.

Limitations of Vision: Multimodal models often hallucinate image content using text cues without actual visual input. However, precise positioning in the physical world cannot be faked.
3-Layer World Models: WorldLabs categorizes world models into renderers (pixels), simulators (geometry/physics), and planners (actions). Video generation alone isn't enough for real-world tasks.
Sim-to-Real Gap: Companies like LimX and AgiBot use simulation platforms for cost-effective trial and error, though virtual parameters can't cover all real-world surprises.
Irreplaceability of Touch: Fei-Fei Li's T-Rex research shows that while vision handles slow global judgments, tactile feedback must operate at high frequencies to provide instant corrections during contact, significantly improving delicate manipulation success rates.

More from Embodied