Embodiment Scaling Laws

• Embodiment Scaling Laws are power-law relations that quantify how performance in embodied systems scales with model size, dataset diversity, compute, and physical morphology.
• They reveal optimal compute allocation between model capacity and data diversity by fitting empirical exponents for tasks like world modeling and behavior cloning.
• These laws extend to biological systems, providing predictions for physical attributes such as locomotion speed and stride frequency across species.

Embodiment scaling laws quantify how performance metrics in embodied systems—ranging from artificial agents to biological organisms—scale with respect to variables such as model capacity, dataset/diversity, task complexity, and physical morphology. Recent advances have extended scaling law frameworks from language and vision models to world models, imitation learning, robot control, and biological embodiment, providing predictive power for resource allocation, generalization, and transfer across morphologies.

1. Mathematical Formulation of Scaling Laws in Embodiment

In embodied pre-training, a scaling law denotes a power-law relationship between an asymptotic performance metric (e.g., cross-entropy loss, mean squared error, or reward) and a “scale” parameter such as model size $N$ (parameters), dataset size $D$ (tokens or trajectories), amount of compute $C$, or the number of distinct embodiments $n$ (Pearce et al., 2024, He et al., 3 Nov 2025, Ai et al., 9 May 2025, Liu et al., 17 Feb 2025).

For transformer-based world models (WM) and behavior cloning (BC) policies, the scaling laws obey: $L_{\rm opt}(C) = \min_{6 N D = C} L(N, D)$

$$N_{\rm opt}(C) \propto C^a,\quad D_{\rm opt}(C) \propto C^b$$

$L_{\rm opt}(C) \approx c_0 C^{-c} + E$

where $C$ is the training FLOPs, $L$ is next-token cross-entropy, and $a$, $b$, $c$, $c_0$, $E$ are empirically fitted constants. For cross-embodiment or multi-morphology settings, the expected generalization metric (reward $J_{\mathrm{test}}$, MSE $E$) scales as a power law in the number of distinct training embodiments $n$ (He et al., 3 Nov 2025, Ai et al., 9 May 2025): $E_{\text{task}}(n) \approx E_0\, n^{-\beta}$

$J_{\mathrm{test}} \approx a\, N^b$

Sub-linear exponents capture diminishing returns as diversity or scale increases.

2. Empirical Laws for Model and Data Scaling in Embodied Agents

Empirical studies using causal-GPT architectures and large human gameplay datasets reveal precise scaling exponents for different architecture/tokenization regimes (Pearce et al., 2024). Table 1 summarizes the relationships:

Experiment	$N_{\rm opt}\propto C^a$	$D_{\rm opt}\propto C^b$
WM-Token (256)	0.49	0.51
WM-Token (540)	0.62	0.37
BC-Token (540)	0.32	0.68
BC-CNN	0.66	0.34

These results indicate:

• For world modeling with high-compression (256 tokens): equal scaling of model and data ($a \approx b \approx 0.5$).
• For tokenizations with higher granularity (540 tokens): increased emphasis on model capacity ($a = 0.62$).
• For behavior cloning with sparse, superclass actions: data scaling is dominant ($a = 0.32$).
• For CNN actions: model scaling regains dominance ($a = 0.66$).

The optimal allocation, given a fixed compute budget $C$, is to set $N \propto C^a$, $D \propto C^{1-a}$ for the relevant architectural regime.

3. Embodiment Diversity and Zero-Shot Generalization

Scaling the diversity of training embodiments, rather than merely data volume, drives generalization across previously unseen bodies. In robot locomotion, the expected zero-shot generalization return $J_{\mathrm{test}}$ grows as a sub-linear power law in the number of training morphologies $N$ (Ai et al., 9 May 2025): $$J_{\mathrm{test}} \sim a \cdot N^b\,,\quad b \approx 0.3$$ For dexterous manipulation, world model prediction error on a held-out morphology decays with training embodiments as (He et al., 3 Nov 2025): $$\begin{aligned} \text{Rigid push:}\quad &\mathrm{MSE} \sim n^{-0.25},\ \text{Plasticine reshaping:}\quad &\mathrm{MSE} \sim n^{-0.47} \end{aligned}$$ This points to diminishing returns, with steeper exponents in tasks requiring richer contact and interaction representations. Empirically, doubling the number of diverse hands or morphologies yields predictable reductions in error or increases in reward, with exponents tuned by the complexity and heterogeneity of the task.

4. Biological Scaling Laws and Physical Embodiment

The scaling of physical size and morphology in animals, microorganisms, and plants is governed by a distinct but complementary set of embodiment scaling laws derived from the invariance of fundamental physical equations under dimensional rescaling (Liu et al., 17 Feb 2025):

• Beating or stride frequency: $f \propto M^{-1/6}$ (empirically, slope: $-0.17 \pm 0.02$)
• Locomotion speed: $V \propto M^{1/6}$ (empirically, slope: $+0.17 \pm 0.03$)
• Leg stiffness in mammals: $K \propto M^{2/3}$ (empirically, slope: $+0.68 \pm 0.04$)
• Frequency vs. length: $f \propto L^{-1/2}$; speed vs. length: $V \propto L^{1/2}$

These laws emerge from coupled fluid–elastic–rigid body PDEs with rescaling of $L$, $T$, $\rho$, $E$, $\nu$, $g$ and are validated across hundreds of species. The same scaling-maps yield predictions even for prehistoric or extinct taxa.

5. Modulators and Practical Implications

The exponents of embodiment scaling laws are heavily modulated by:

• Tokenizer compression: Increasing tokens per image (low compression) boosts the model-scaling exponent, e.g., $a$ shifts from 0.49 (256 tokens) to 0.62 (540 tokens) (Pearce et al., 2024).
• Loss sparsity and granularity: Sparse action losses in BC slow model saturation, causing $a$ to drop ($\sim 0.32$). Coarser loss increases data scaling dominance.
• Architecture compute-density: Feeding more tokens per action increases per-step compute, affecting the "frontier" at which further scaling plateaus. Re-encoding actions as dense embeddings can re-balance scaling back toward model size.
• Dataset diversity: Large, diverse datasets (e.g., 8.6 years of human play) prevent reuse and clipping of power laws, ensuring reliable empirical exponents.
• Task complexity and morphology heterogeneity: Tasks with rich local interactions (e.g., deformable manipulation) benefit disproportionately from diversity scaling (steeper $\beta$ exponent).

A plausible implication is that optimal scaling for a given resource budget requires matching architecture, data, task encoding, and desired generalization profile. For policies intended for broad real-world deployment, heavy investment in embodiment diversity (not just more data per morphology) offers the most robust zero-shot performance.

6. Synthesis and Applications

Embodiment scaling laws unify trends observed in deep RL, imitation learning, world modeling, and biological physics. They explain:

• How to allocate compute between model and data for embodied agents.
• Why generalist policies exposed to orders of magnitude more morphologies outperform single-morphology or pure "big data" baselines in transfer settings.
• The evolutionary logic in animal and plant biomechanics, revealing that power laws in frequency, speed, and stiffness are a mathematical consequence of geometric scaling and PDE invariance.

These laws provide quantitative recipes that practitioners can use for architecture selection, resource planning, and extrapolation. For robotics, leveraging these principles enables the training of universal controllers and world models capable of generalizing across unseen morphologies and environments, with broad application to adaptive control, modular robots, and automated co-design of hardware and policy.

Domain	Scaling Law (form)	Exponent (empirical)
Agent model vs. data	$N_{\rm opt} \propto C^a$	$a=0.32$--$0.66$
Embodiment diversity	$J \propto N^b$ or $E \propto N^{-\beta}$	$b=0.28$--$0.47$
Biology (animals)	$f \propto M^{-1/6}$, $V \propto M^{1/6}$	$-0.17$, $+0.17$

These results substantiate embodiment scaling as a foundational concept in both artificial and biological embodiment, linking model, data, and morphological diversity to performance generalization and efficiency.