The evaluation applies eight kinds of observation corruption at exactly one inference step per episode across five . Failures depend strongly on timing: at vulnerable steps, every model loses more than 50 percentage points of success under at least one perturbation.
Longer action execution chunks amplify the damage because a robot continues following a bad predicted sequence. CARE compares consecutive action predictions and temporarily shortens execution when they disagree, improving robustness with low overhead while largely preserving clean-task performance.