Robotics & manufacturingResearch findings
ReasonBreak · Vulnerability Analysis of Reasoning-Enabled Autonomous Driving Models
Original published 2026.05.28 06:21 KSTAdded to Blackdeck 2026.11.04 01:42 KST
What has changed?
First systematic black-box study on vulnerability to realistic input corruptions in reasoning-enabled VLA models
Key points
- Reasoning-enabled VLA models are highly vulnerable to realistic input perturbations
- Attack success rates on reasoning and trajectory manipulation reach up to 89% and 72% in closed-loop simulation
- Proposes a new evaluation framework and benchmark for reasoning-trajectory interaction
Connect it to your work
Introduces a reasoning-aware evaluation framework capturing semantic and structural aspects of reasoning
What remains unproven
The abstract reports simulation results; real-world performance is not established by this input.
Originals used for this note
Links open the original publisher in a new tab.