General Physical Intelligence

• General Physical Intelligence is a multidisciplinary framework that integrates robotics, materials science, AI, and cognitive systems to enable real-world perception, reasoning, and action.
• It employs methodologies from embodied AI, multiscale materials modeling, and advanced simulation to quantify both micro-level task solving and macro-level planning.
• GPI research drives innovations in robotics, advanced manufacturing, and governance through multimodal foundation models and causal scientific inference.

General Physical Intelligence (GPI) represents a multifaceted concept at the intersection of robotics, materials science, artificial intelligence, and cognitive systems. GPI encompasses the ability of systems—natural or artificial—to perceive, reason, and act within the physical world, combining high-level cognition with physically grounded capabilities. The concept has evolved through several disciplinary threads including embodied AI, ferroelectric materials modeling, multimodal foundation models, governance frameworks for physical AI, and general theories of intelligence grounded in information processing. Recent research provides theoretical, methodological, and empirical grounding for GPI, with particular emphasis on its role in robotics, advanced manufacturing, and scientific inference.

1. Theoretical Foundations: Intelligence as Information Processing

GPI is distinguished from both narrow physical reasoning and disembodied cognitive intelligence by its grounding in real-world interaction (Hochberg, 2023). The "Theory of Intelligences" (TIS) frames intelligence as a universal process of uncertainty reduction towards goals—a calculus composed of differentiation (decomposition of goals into subgoals), correlation (matching experience to priors), and integration (synthesizing solutions into broader plans). Thus, GPI systems not only solve immediate sensory-motor tasks but also plan and optimize sequences of actions. Macroscopic indices quantify both "solving" at the micro-level and "planning" at the macro-level, with formal expressions such as:

$$\mathbf{I}_{x,y} = \mathbf{U}^{R}_{x,y} \cdot (\alpha + \beta \mathcal{A}^G_{x,y})$$

where $\mathbf{U}^{R}_{x,y}$ represents normalized solving and $\mathcal{A}^G_{x,y}$ quantifies planning efficiency. Goal difficulty is expressed as a ratio between intrinsic complexity and system ability, e.g., $D_y = \mathcal{C}_{g,y} / Q_{x,y}$. This formalism applies equally to physical, biological, and artificial systems.

2. Multiscale Modeling: From Materials to Robotics

Original foundational work on GPI in materials science developed a microscopic model for glycinium phosphite ferroelectric crystals, integrating proton ordering with piezoelectric coupling (Zachek et al., 2017). The Hamiltonian for the system is partitioned into seed energy $U_{seed}$ (elastic and piezoelectric terms), short-range proton interactions, and long-range dipole-dipole coupling, expanded linearly in strain variables ($\varepsilon_i$). The model robustly predicts components of the polarization vector and the dielectric permittivity tensor under transverse external fields ($E_1$, $E_3$):

$$P_1 = e_{14}^0 \varepsilon_4 + e_{16}^0 \varepsilon_6 + \chi_{11}^{\varepsilon 0} E_1 + \chi_{31}^{\varepsilon 0} E_3 + \frac{1}{2v}[\mu^{x}_{13}(\eta_1 - \eta_3) - \mu^{x}_{24}(\eta_2 - \eta_4)]$$

Thermodynamic characteristics (e.g., phase transition temperature shifts) are derived by minimizing the free energy with respect to strain, capturing the interplay between external stimuli, microscopic ordering, and macroscopic response. Such approaches embody the predictive and control aspects central to GPI.

3. Embodied Intelligence: Simulators, World Models, and Foundation Architectures

Recent surveys delineate GPI in robotics in terms of embodied intelligence, emphasizing the integration of multimodal perception, closed-loop control, and world-model simulation (Long et al., 1 Jul 2025). Physical simulators (MuJoCo, Isaac Gym, Webots) provide safe, high-fidelity platforms for agent training; world models construct internal generative representations for predictive planning and adaptation. The synergy enables agents to learn transferable behaviors—critical for sim-to-real deployment. Embodied intelligence is formalized through policies maximizing cumulative reward, subject to physical constraints:

$$s_{t+1} = f(s_t, a_t) \qquad \pi(s) = \underset{a}{\text{argmax}}\, \mathbb{E}[R | s, a]$$

Systematic benchmarking, hierarchical architectures, and integration of cross-modal reasoning drive progress toward generalizable, robust GPI.

4. Multimodal Foundation Models: Vision Language Action Integration

In agile manufacturing and advanced robotics, foundation models—particularly Vision Language Action (VLA) frameworks—are central to practical GPI implementation (Kanta et al., 16 Aug 2025). These models fuse visual (DINOv2, SigLIP), linguistic (LLaMA), haptic, proprioceptive, and machine-state information into unified latent spaces for contextual reasoning and action. Examples such as RT-2-GPI, PaLM-E-GPI, VIMA-GPI, and Gato-GPI are assessed through comparative ablation studies:

• RT-2-GPI: Top success rate (93.3%), high generalization (0.89), deep affordance integration
• VIMA-GPI: 88.3% success, moderate generalization, strong object memory but less multimodal robustness
• Gato-GPI: 83.3% success, slower inference, weaker adaptation to novel tasks

Task evaluation is formalized using policy definitions and loss functions:

$$a_t \sim \pi_\theta(a_t | \tau(o_{\leq t}), \hat{s}_{t:t+H}, \mathcal{G}, \mathcal{E})$$

$\mathcal{L}_{GPI} = \mathcal{L}_{action} + \lambda_1 \mathcal{L}_{world model} + \lambda_2 \mathcal{L}_{goal\mathchar`-success}$

Deterministic fusion of asynchronous data streams, sim2real transfer protocols, and uncertainty quantification (conformal prediction, dropout) are recognized as decisive for industrial readiness.

5. Measurement, Benchmarking, and Evolutionary Learning

GPI as realized in humanoid robotic systems is operationalized via multidimensional performance indices (e.g., the g⁺ metric), derived from the O*NET taxonomy of human work (Gildert et al., 2023). The work fingerprint is a 120-dimensional vector, and g⁺ is computed as:

$$g^+ = \left( \sum_{i=1}^{120} s_i \right) \left( \frac{100}{267.3} \right)$$

Evolutionary learning workflows incrementally improve robot capabilities: data collection from human experts, decomposition into Instruction Set tasks, iterative teleoperation, and autonomous execution yield measurable increases in g⁺ over time. Historical measurements indicate a current g⁺ of 78.2 (teleoperation) and 73.7 (autonomous), with trajectory projections suggesting human-level coverage within 60 months given steady improvement.

6. Governance and Societal Integration

GPI deployment at scale necessitates comprehensive governance frameworks to address existential risks, data organization, disciplinary bottlenecks (Cannikin Law), and social acceptance (Li et al., 2023). Governance models involve multidimensional knowledge graphs for entity-relationship tracking ($K = (V, E)$), risk scoring functions:

$$R(\text{PAI}) = \alpha R_{existence} + \beta R_{data} + \gamma R_{disciplinary} + \delta R_{social}$$

Challenges include regulatory heterogeneity, interdisciplinary integration, ethical concerns (unemployment, privacy), and safe transfer from controlled to public environments. The review points to the necessity of holistic, adaptive governance architectures for responsible GPI integration.

7. Scientific Inference and Causal Reasoning

GPI also encompasses the capability for causal scientific inference using AI-based models. Advances in causal discovery (e.g., LiNGAM) enable systems to move beyond association towards mechanistic understanding via deliberate interventions (Singh et al., 2023). Directed Acyclic Graphs model variable relationships; causality is operationalized using the do-operator:

$P(Y | do(X)) = \sum_z P(Y | X, z) P(z)$

This approach enables reliable prediction and manipulation of physical systems, supporting the development of GPI systems with robust experimental and reasoning capabilities.

8. Future Directions and Open Challenges

Key open challenges for GPI include the establishment of "contact-rich" benchmarks linking haptic/proprioceptive data with explicit affordances (Kanta et al., 16 Aug 2025), hierarchical planning integrated with temporally aligned feedback control, compressing high-capacity models to meet real-time constraints, and enhancing interpretability for safety-critical deployment. From a theoretical perspective, evolutionary signatures in intelligence traits, proxy-based extensions, and multi-level integration remain active research themes (Hochberg, 2023). Domain-agnostic frameworks (e.g., GenAI-Powered Inference) exemplify the application of GPI methodologies to unstructured, high-dimensional data (Imai et al., 5 Jul 2025), further expanding the scope of physical intelligence beyond traditional robotics.

Summary Table: Core Dimensions and Representative Models

Dimension	Representative Approach	Key Metrics or Components
Microscopic Modeling	GPI crystal pseudospin-lattice Hamiltonian	$P_i$, $\chi^{\varepsilon}_{ij}$, $T_c$
Embodied Intelligence	Simulators (Isaac Gym, MuJoCo), World Models	$s_{t+1}=f(s_t,a_t)$, $\pi(s)$
Foundation Models	VLA frameworks (RT-2-GPI, PaLM-E-GPI)	Success rate, generalization score
Benchmarking	Work fingerprint, g⁺ metric, O*NET taxonomy	$g^+ = (\sum s_i)\cdot(100/267.3)$
Governance	Knowledge graphs, risk scoring functions	$K = (V, E)$, $R(PAI)$
Causal Inference	DAGs, do-operator, symbol regression	$P(Y|do(X))$, LiNGAM

General Physical Intelligence thus emerges as an overview of physical modeling, embodied cognition, causal scientific reasoning, immersive multimodal architectures, and rigorous governance. Its realization depends on the tight integration of perception, planning, action, and adaptation across sensors, physical agents, and informational substrates, with performance and safety measured by domain-specific, empirically validated benchmarks.