Ngan Le

Vision-Language-Action (VLA) models adapt pretrained Vision-Language Models (VLMs) for robotic control but often bind instructions to incorrect visual evidence. Single-object tests show that state-of-the-art policies grasp the visible object even when the requested referent is absent. We ask whether explicit grounding can improve instruction-scene alignment under visual distraction and distribution shift without distractor-rich demonstrations or auxiliary perceptual objectives. We propose Distraction-Robust Action via Geometric and Object-ceNtric Grounding (DRAGON-VLA), an integrated perception-control system. Its two-stage object-centric module uses an off-the-shelf VLM to select segmentation proposals in a base view and match them semantically in a wrist view. Its geometric module maps selected regions to masked monocular relative depth, preserving shape and layout while suppressing appearance, background, and distractors. These inputs condition a pretrained VLA fine-tuned only on clean, single-object demonstrations. DRAGON-VLA offers a task-aware, calibration-free interface combining base-view anchoring, cross-view matching, and depth grounding without additional high-level grounding supervision. On a real UR10e tabletop setup, it outperforms VLA baselines under distractors, absent referents, background shifts, and held-out categories. With seven distractors, grounded policies retain around 80% task success while baselines fall below 10%; they also achieve approximately 95% absent-target rejection. Ablations identify object-centric filtering as the main source of gains, with masked relative depth providing its clearest benefit on unseen objects. These results support explicit grounding for referentially reliable VLA manipulation. Project page: https://uark-aicv.github.io/DRAGON_VLA/.