Vision language action models act despite sensing danger
Beyond Behaviour: A Stimulus-Response Safety Evaluation of Vision-Language-Action Models
Artificial Intelligence
Summary
Some AI models can see and understand both pictures and words and act on them, like a robot might. This paper looks at whether these models react to dangerous things the way humans do — by stopping or changing their actions to stay safe. The researchers found that even when the models notice danger inside, they usually keep doing their original tasks. This means the models don’t naturally fight or avoid unsafe situations the way people might.
What this means in practice
- •For robotics engineers: Evaluate robot control systems to identify whether internal hazard detection influences action decisions to improve safety measures.
- •For autonomous vehicle developers: Assess and improve vision-language-action model responses to visible dangers in traffic scenarios to reduce accident risks.
Authors
Vy Vo, Dinh Phung
Abstract
Humans exhibit biological sensitivity to safety-critical stimuli that can trigger defensive action. Across three models and two video benchmarks, we investigate whether similar responses emerge in vision-language-action models by examining their internal representations and action tendencies in response to visible hazards and harmful consequences. Our findings suggest that task execution may remain the default response even when danger is internally registered.