EgoBlind-RA: Towards Safer Egocentric Assistive AI for Blind Users via Risk-Adaptive Routing

Blind and low-vision (BLV) users depend on visual assistance systems to navigate an unpredictable physical world, yet existing multimodal systems apply a uniform inference policy irrespective of query urgency. We present EgoBlind-RA, a risk-adaptive framework that classifies egocentric query-video pairs as urgent or non-urgent via a lightweight CLIP-based classifier and routes each query to a response calibrated to its safety stakes. Our classifier achieves 0.905 ROC-AUC on the held-out test set. We compare two LoRA-fine-tuned architectures of Kimi-VL-A3B-Instruct on the EgoBlind benchmark: a bifurcated design with separate urgent and non-urgent adapters, and a unified prompt-tag-conditioned design. Urgency-conditioned SFT against carefully selected reference targets substantially improves over a uniform-policy baseline in both. Surprisingly, DPO regresses SFT in the unified setting whilst proving productive in the bifurcated setting, where the best result mixes an SFT urgent adapter with a DPO non-urgent adapter. The unified adapter is also robust to noisy CLIP routing at deployment, matching oracle composite loss despite 21.5% test-time tag flips. Together, these results suggest the value of preference-pair-based alignment and the cost of classifier noise both depend sharply on architectural regime.

MIT MMAI