How machine learning researchers build a better world model for robot decisions

Learning from unexpected outcomes

Randall Balestriero of Brown University and AMI Labs, Kerda Varaku, and Yann LeCun of NYU and AMI Labs present the approach in a pre-print paper titled “Counterfactual World Models: Natura Non Facit Saltus.” The Latin phrase refers to Gottfried Leibniz’s axiom that nature makes no leaps. The method estimates what might have happened under an alternative action by using the difference between a model’s prediction and the real outcome. Balestriero said this could support better future predictions, debugging, and reasoning at test time.

Applying the correction

The researchers say the method can work with any world model that predicts an average next state for an action, though JEPA models are especially suitable for adaptation. A typical JEPA model averages possible outcomes and cannot, on its own, anticipate events that did not occur.

In Balestriero’s boat example, the model predicted the result of steering left, but the boat ended up two meters farther downstream. The researchers treat that difference as the current’s effect, then apply it to the model’s prediction for steering right to estimate the alternative outcome.

Results in simulated tasks

The paper says the error in estimating the disturbance’s effect decreases with the square of the time step. The authors also say no method using the same ordinary data can achieve a better rate in the worst case.

When tested in simulations, selecting actions with this approach reached the goal in 35 to 69 percent of events. A planner that knew the true simulator but disregarded the observed outcome succeeded in 14 to 23 percent; after two further attempts on the same event, the counterfactual method’s success rose to 51 to 97 percent.

Why prediction is not enough

Balestriero said real-world robots, cars, and drones encounter surprises they could not anticipate. After a failure, he argued, it matters whether a different action would have succeeded in the same circumstances, helping distinguish a poor decision from bad luck.

The researchers say hidden disturbances cannot generally be learned from ordinary data alone, because they can be relabeled for different actions without changing predictions. Their proposed correction instead relies on the mismatch between prediction and outcome, along with the gradual accumulation and addition of forces in physical systems.