The Embodied Computation Trilemma: Thermodynamic Boundaries, Phase-Space Collapse, and Deterministic Control in Vision-Language-Action Models
Year: 2026
Type: preprint
Abstract: Contemporary physical artificial intelligence (Physical AI) and vision-language-action (VLA) foundation models rest upon an unproven foundational axiom: that dynamic interaction within continuous physical space-time can be asymptotically resolved by scaling discrete floating-point operations (FLOPs) and multimodal autoregressive tokens. In this work, we formalize the fundamental limits of this paradigm as The Embodied Computation Trilemma, establishing a rigorous mathematical and thermodynamic bound connecting control efficiency ($E$), hardware/compute cost ($C$), and physical safety bounds ($S$): $$\frac{E \cdot S}{C} \le k_{\text{embodied}}$$ Grounded in micro-scale Landauer dissipation and macro-scale second-law thermodynamic constraints, we prove that unconstrained optimization toward maximizing efficiency ($E \to 1$) while stripping away physical safety margins ($S \to 0$) inevitably induces Phase-Space Divergence and Material Constitutive Inadmissibility when confronting discontinuous contact boundaries and friction bifurcations in the real world. To resolve this epistemic deadlock, we introduce a deterministic, zero-heap C99 fault-tolerant state estimator as a zero-stack safety layer, ensuring bounded momentum reacquisition for hard real-time embedded systems. This work bridges abstract algorithmic assets with physical reality, backed by our comprehensive 20-part Physical Intelligence (PI) matrix.
Author: Ye, Qiang(Peter)
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225909215
