Insufficiency of Reactive Inverse Models for Obstacle-Constrained Rope Manipulation

Establish that an inverse dynamics controller learned from image observations of rope manipulation, which computes actions reactively from consecutive observations without explicit planning, cannot by itself plan movements that involve fixed obstacles in the rope manipulation environment.

Background

The paper introduces Visual Planning and Acting (VPA), which combines a Causal InfoGAN-based visual planner with a learned inverse dynamics controller to manipulate deformable objects such as rope. In the real robot experiments with a PR2, the environment includes fixed obstacles that the rope cannot pass through, making planning nontrivial.

In this setting, the authors explicitly conjecture that a purely reactive inverse dynamics model—trained to map pairs of consecutive rope images to actions—will be insufficient to plan movements around obstacles. They then empirically compare this baseline against VPA, providing evidence that planning is necessary to navigate around obstacles, but they do not offer a formal proof of the conjecture.