Disentangle the effects of individual RoboTwin visual perturbations

Determine the individual effects of lighting changes, background-texture changes, clutter, and table-height variations on the performance of the K1 perception-augmented robot-use agent in RoboTwin bimanual manipulation tasks.

Background

The RoboTwin transfer experiment evaluates Gemini with the K1 harness under Easy and Hard conditions. The Hard condition simultaneously varies clutter, illumination, textures, and table height, and K1 exhibits a smaller Easy-to-Hard performance drop than published VLA baselines. Although this pattern suggests that calibrated spatial measurements may improve robustness to appearance changes, the experimental design combines several perturbations, preventing attribution of the observed effect to any single factor. The unresolved problem is therefore to isolate and quantify the contribution of each perturbation independently.