Why Physical AI needs imagination. The math of object permanence

When a human operator drops a wrench behind a toolbox, they don't assume the wrench has ceased to exist. They instinctively maintain a mental model of the object's continued presence, location, and geometry despite total visual occlusion. In developmental psychology, this is known as object permanence.

In the realm of robotics, a surprising number of advanced Vision-Language-Action (VLA) models completely lack this capability. They are highly performant reactive engines that operate frame-by-frame. If a forklift drives in front of a staging pallet, or if a dense part obscures another in a bin-picking operation, these systems frequently fail. To them, the occluded object has been erased from reality.

The status quo in much of the industry has been to build "thicker" perception networks. The assumption is that by scaling parameters, the neural model will implicitly learn to interpolate missing data. Some implementations go a step further, relying on rudimentary temporal smoothing or short-horizon LSTM architectures to "remember" the last few frames.

However, simple persistence memory or implicit feature-space memorization scales poorly in cluttered, dynamic environments. The physical world contains long occlusions, moving actors, and complex geometric intersections. A reactive system that simply holds a stale 3D coordinate until an object reappears will invariably collide with the environment if the object itself is moving or if the robot needs to interact with the hidden geometry.

Intelligence in physical AI is not merely the ability to perceive. It is the ability to maintain a mathematically rigorous hallucination of the unobserved.

To solve occlusion, we must move from a paradigm of perception (What do I see?) to a paradigm of state estimation (What is the true state of the world?).

Mathematically, this demands a probabilistic approach, conceptualizing the environment as a hidden Markov model where the true state $x_t$ is unobserved, and we are given a sequence of noisy observations $z_{1:t}$. The objective is to maintain a belief distribution over $x_t$, expressed as $p(x_t | z_{1:t})$.

At Xolver, addressing this takes the form of an active World Model. Instead of merely responding to pixels, the system continuously predicts the forward evolution of the scene based on physical priors. This process is governed by the recursive Bayesian update equation:

This equation contains two critical phases. First is the *Prediction Step* (the integral). Here, the model applies transition dynamics $p(x_t | x_{t-1}, u_{t-1})$, effectively hallucinating how the world—including both the robot and dynamic actors—should evolve over the time step. Second is the *Correction Step*, where the incoming visual observation $z_t$ grounds the hallucination, updating the probability distribution.

Understanding the mathematics of state estimation is only half the problem. The harder engineering challenge is integrating this probability distribution into a deterministic control loop.

If a robot is tasked with moving through a cluttered warehouse aisle, and a worker steps behind a pallet stack, the robot’s World Model now holds a probability distribution of the worker's location behind the occlusion. But the underlying motor-driver cannot be given a probability. It demands a definite command.

This is where the Xolver architecture of Bounded Execution is vital. Our foundation model proposes an intent, incorporating its probabilistic estimate of the occluded world. The Deterministic Enforcement Layer then takes this proposal and evaluates it against spatial and kinematic constraints.

Crucially, the constraint boundaries are dynamically expanded based on the variance (uncertainty) of the unobserved state. If the World Model is highly uncertain about the position of a hidden object, the keep-out zone inflates, forcing the path planner to take a wider berth or reduce velocity. The system operates confidently up to the very margin of its mathematical uncertainty.

A robot that cannot imagine what it cannot see is doomed to fail in a complex physical environment. By grounding physical AI not just in perception but in recursive probabilistic modeling, we enable machines to hold object permanence as a mathematical truth.

At Xolver, we do not view occlusion as an edge case or a perception failure. It is the fundamental state of the real world. By building World Models that predict the unseen and control loops that respect the geometry of uncertainty, we are moving beyond reactive operations and toward genuine spatial intelligence.