Vision-language-action (VLA) models assume cameras stay fixed; if bumped, performance drops even though the task is unchanged.
Homeward transforms new views back to the training view without training: it estimates camera position and scene depth from a single reference image, reprojects the scene, and renders the robot arm using known joint angles.
On 1,599 camera positions for LIBERO-Plus, it raised success from 62.5% to 77.6%, enabling robust deployment without retraining or calibration.
