Accepted paper ·
"Observe Before You Alert" accepted to CoRL 2026
VLAlert casts driver alerting as a tri-action policy — SILENT, OBSERVE, ALERT — and releases VLAlert-Bench: 13,534 videos and 192,892 labelled one-second ticks from six driving-event datasets.
Observe Before You Alert: Adaptive Driver Alerting with Vision–Language Models (Yuhang Wang, Lingyao Li, Hao Zhou) has been accepted to the Conference on Robot Learning (CoRL 2026). The OBSERVE action lets the system gather evidence before it warns; on VLAlert-Bench validation the policy reaches a DAUS of 0.4878 versus 0.4752 for Open-BADAS, and on 221 held-out ADAS-TO-Critic clips it raises R@5s from 74.2 % to 88.7 %. Demos, tables and the benchmark’s access terms are on the project page.