The cost of thinking. Why low-level control must be dumb

As robotics foundation models scale from millions to billions of parameters, they demonstrate unprecedented semantic reasoning. A Vision-Language-Action (VLA) model can interpret messy natural language instructions, evaluate complex camera feeds, and infer intricate spatial trajectories. But this intelligence comes with a steep physical tax: latency.

In modern transformer architectures, autoregressive token generation or multi-step diffusion sampling typically requires anywhere from 50 to 200 milliseconds per inference cycle. In the digital world, a 100-millisecond delay while rendering text or generating an image is completely imperceptible. But in physical mechanics, 100 milliseconds is an eternity.

Consider a 50kg industrial manipulator moving at a moderate velocity of 1.5 meters per second. If the control system pauses for 100 milliseconds to wait for the next neural network forward pass, the robot travels 15 centimeters in a completely ungoverned state. During that latency window, physical momentum, unexpected friction, or sudden human movement can turn a safe trajectory into a high-energy collision. When intelligence is slow, controlling motor torques directly with a neural network is an invitation to failure.

To understand why end-to-end neural control fails at scale, we must look to classical frequency-domain analysis and the dynamics of continuous feedback control.

Let a physical robot be governed by the continuous non-linear state space equation:

where $x(t) \in \mathbb{R}^n$ represents the state vector (joint angles, angular velocities, and torques) and $u(t) \in \mathbb{R}^m$ is the control input. When the control decision $u(t)$ is generated by a deep neural network subject to computational latency $\tau_{\text{think}}$, the system transforms from an ordinary differential equation into a Delay Differential Equation (DDE):

In the frequency domain, a pure time delay $\tau_{\text{think}}$ acts as an exponential operator $e^{-s \tau_{\text{think}}}$ in the open-loop transfer function $G(s)H(s)$. When evaluated at operational frequency $\omega$, this delay introduces an uncompensated phase drop:

This phase lag directly erodes the system's Phase Margin ($\text{PM}$):

As model size increases and $\tau_{\text{think}}$ grows, the phase margin rapidly drops toward zero. The consequence is not merely sluggish tracking, but destructive limit cycles, resonant chatter, and violent closed-loop instability. Furthermore, according to the Nyquist-Shannon sampling theorem, to stabilize physical dynamics with structural resonance at frequency $\omega_{\text{res}}$, the closed-loop control frequency must satisfy $f_{\text{control}} \ge 2 f_{\text{res}}$. For industrial joint drives, structural resonance demands control frequencies of at least $1000\text{ Hz}$ ($1\text{ ms}$ execution budgets). An AI inference loop running at $10\text{ Hz}$ or $20\text{ Hz}$ violates this mathematical bound by two orders of magnitude.

Nature solved this exact computational trade-off millions of years ago.

The mammalian central nervous system does not run muscle actuators directly from the cerebral cortex. The cortex is a slow, highly deliberative probabilistic engine. Processing visual inputs, evaluating environment semantics, and deciding on a motor plan requires between 100 and 300 milliseconds of neural processing time.

If biological motor control relied exclusively on the cortex, touching a scalding hot surface would result in severe tissue destruction before the brain could transmit a single corrective command. To survive, biological evolution developed frequency separation.

The monosynaptic stretch reflex arc in the spinal cord operates locally at microsecond timescales. The spinal cord is fundamentally "dumb"—it has zero semantic understanding of what an object is. However, it possesses instant access to proprioceptive muscle spindle feedback, executing deterministic reflex actions ($1000\text{ Hz}$) that maintain postural stability and enforce immediate safety boundaries. The cortex sets high-level target intent, while the spinal cord governs continuous physical momentum.

Translating this biological principle into physical AI requires an architectural separation of concerns: decoupling slow probabilistic reasoning from fast deterministic enforcement.

Rather than attempting to force deep foundation models to run at $1000\text{ Hz}$, physical AI platforms must adopt a multi-rate control hierarchy:

Mathematically, for any timestamp $t \in [t_k, t_{k+1}]$ between model inference steps, the $1000\text{ Hz}$ runtime solves the instantaneous optimization:

If the foundation model stalls, drops a frame, or encounters an inference lag, the fast deterministic layer does not crash or pause in an ungoverned state. It seamlessly smoothly decelerates the physical system along the verified manifold $P_k(t)$ to a zero-velocity equilibrium state. The robot remains safe regardless of the model's computational delays.

High-level intelligence should be deep, deliberative, and semantic. But low-level control must remain ultra-fast, deterministic, and fundamentally dumb. Attempting to blur this boundary by asking foundation models to directly control microsecond joint torques misunderstands the physics of feedback control.

At Xolver, our Edge Runtime is built on this immutable law of physical systems. By isolating heavy model reasoning from microsecond deterministic loops, we ensure that as AI models grow larger and smarter, hardware execution remains unconditionally fast, stable, and safe.