World Models for Robots: Why Conditioning on Friction Matters More Than the Model Itself

In a July 2026 piece, The Robot Report stresses that for world models to work on a real robot rather than in a simulator, one separate and often underrated step is needed: conditioning the model on the physics of its environment, and above all on friction.

What is a world model for a robot

A world model is a trained model that predicts how the environment will change in response to a robot's action: whether an object will shift, whether the grip will hold, where a part will roll. The robot effectively gets an internal "simulator" and plays out the consequences of an action there before performing it in reality. That is why world models have become one of the hot topics in robotics: the model "imagines" the outcome instead of trying dozens of movements on live hardware, and every such trial on a real robot costs time and risks damage.

The problem is that this "imagination" is only as useful as it is faithful to real physics. It is exactly at this boundary between the trained model and real contact that the main deployment difficulty arises.

Why conditioning is underrated

Conditioning means fitting a trained model to the specific conditions in which the robot operates: surface properties, the weight and shape of objects, and the characteristics of the gripper. The Robot Report notes that this step gets less attention than the architecture and training of the models themselves, even though it determines whether quality carries over from simulation to a real robot. A model can perform brilliantly in an idealized environment and fall apart on a real tabletop if it hasn't been adapted to what the robot actually touches.

"Deploying world models in real robotic systems requires a step that gets less attention than the models themselves: conditioning,"

The Robot Report writes.

What does friction have to do with it?

Friction determines whether a robot holds an object or drops it, whether a part slips in the gripper, and whether there is enough force to push an object to the right spot. If a world model was trained on simplified physics, its contact predictions diverge from reality first and foremost on friction, the most finicky and worst-modeled parameter of the interaction. A small error in the friction coefficient turns into a failed grasp or a botched assembly. So, by the logic of The Robot Report's piece, correctly tuning friction is the key to getting a world model to produce predictions you can rely on when controlling a live robot.

What does this change for engineers?

For teams transferring trained policies from simulation to hardware, The Robot Report's conclusion is practical: effort should go not only into the world model architecture itself but also into the conditioning stage. The gap between simulation and reality, the so-called sim-to-real gap, has been the main barrier to robotic manipulation for years, and friction sits at its center. A model carefully tuned to real friction coefficients and gripper parameters makes noticeably fewer errors on contact than the same model out of the box. This shifts the focus of development: quality on a live robot depends not only on how smart the model is, but also on how carefully it has been grounded in the specific physics.

What this means

World models bring robots closer to "thinking through" an action before moving. But between a trained model and a working robot stands an unglamorous but decisive step: conditioning on real physics. As long as this tuning, and friction modeling in particular, remains the weak link, impressive results from simulation will keep stumbling over hardware.