Academic boundaries: three melting lines
VLA x world models x AGI: the 2026 cross-citation network
Line one: embodied AI vs classical robot control
Classical control is model-plus-planner: build dynamics then solve optimal control; embodied AI is data-plus-policy: VLAs map vision and language directly to actions (a_t = pi(s_t, l)). The first VLA survey in IEEE TNNLS 2026 splits research into three lines: component-level (vision/language/action encoders), policy-level (low-rate high-level planning plus high-rate low-level control), and planner-level (long-horizon task decomposition). The term VLA was coined by RT-2.
Line two: VLA vs world models
VLAs' soft spot is physics-blindness: poor continuous dynamics, unverified plan feasibility, data scarcity. World models patch all three as future predictors: the Tongji/UESTC survey (Jan 2026) gives four paradigms - world planner (search plans in imagination), world action model (fold prediction into policy), world synthesizer (generate training data), world simulator (interactive sandbox). UBTECH Thinker-WM and NVIDIA Cosmos 3 are industrial implementations of the synthesizer-plus-simulator route.
Line three: embodied AI vs AGI
Both 2026 surveys (IEEE TNNLS and SSRN's 'Large Model Empowered Embodied AI') write embodied AI as a cornerstone/viable path to AGI, on verifiability: textual intelligence can hallucinate, physical intelligence gets falsified on the spot by gravity and friction. Conversely AGI's reasoning (long-horizon planning, causality, counterfactuals) is exactly what embodied AI lacks. Hence 2026 papers show a two-way convergence: AGI benchmarks add physical manipulation items, embodied benchmarks add language-reasoning items. The boundary is not vanishing - it is becoming an interface.
| Concept | Core question | Representative 2026 work |
|---|---|---|
| VLA | End-to-end language+vision to action | RT-2 line, GR00T N1, Helix |
| World models | Predict continuous futures, verify plans | Thinker-WM, Cosmos 3/Edge |
| AGI | Cross-domain general reasoning | Physical items entering AGI benchmarks |