"Probing Frozen Visual Representations for Safety-Relevant Structure in MiniGrid LavaGap Navigation: A Geometric Comparaison of DINOv3 and ResNet-50"

Robots operating in the physical world must recognize when they are approaching danger. This study asks whether frozen pretrained visual encoders already encode safety-relevant structure without any safety supervision. In our MiniGrid LavaGap navigation environment, agent states were labeled safe, near-boundary, or unsafe, and a self-supervised encoder (DINOv3) was compared with a supervised ImageNet encoder (ResNet-50). Embeddings were analyzed geometrically and probed with lightweight classifiers, and generalization was tested on withheld layouts without retraining.

Two findings emerge. First, detecting unsafe states requires no learning representation: an untrained network identifies lava about as well as either pretrained encoder. Second, the meaningful difference lies between safe and near-boundary states. On withheld layouts, frozen DINOv3 separates these classes (balanced accuracy 1.000, AUROC 0.975), whereas frozen ResNet-50 fails to generalize (balanced accuracy 0.528, AUROC 0.750). A leaky evaluation protocol had concealed this divergence. It appeared only under held-out-layout testing.

These results suggest that self-supervised features encode the agent's proximity to hazard boundaries, whereas ImageNet-supervised features do not capture it reliably. They also show that evaluation design can hide, not only inflate, differences between representations, au cautious for future work on representation-based safety monitoring in embodied AI.