Researchers asked whether a model that predicts how a deformable rope will move also helps a robot choose the best actions. They built Geometry‑Pointer‑Flow (GPF), a policy that splits choosing where to grasp the rope from deciding how to move it.

A pure geometry version of GPF cut the one‑step error of the rope shape by 32.8 % compared with a diffusion policy while using 18.7 × fewer parameters and running 193 × faster. The authors then compared three ways of providing physical context to GPF: true rope parameters, estimated parameters, and a history‑based representation learned from past interactions.

In 470 paired simulation trials, the hist‑based representation (Hist‑FiLM) gave the best execution. It lowered the error from 9.257 mm (baseline) to 8.336 mm, a 0.921 mm improvement or 9.9 % relative gain.

The study shows that a context that improves prediction does not automatically improve action selection. In simulation, the estimated context predicted rope behavior better, but with no control benefit. On real hardware, the matched Histogram context gave the lowest prediction error, while Hist‑FiLM achieved the best closed‑loop performance. Which additional information the history carries remains unknown.

Because real‑world trials cannot start from identical states, the experiments matched only by target task and initial conditions, limiting the precision of zero‑shot transfer analysis.