Does VLA Even Know the Basics? Measuring Commonsense and World Knowledge Retention in Vision-Language-Action Models

Vision-Language-Action (VLA) models, used in robotic manipulation, show promise but struggle with complex tasks and richer semantic categories.

Researchers have been studying VLA models, which are typically obtained by fine-tuning powerful pretrained Vision-Language Models (VLMs) on robotics data^[1]. A recent study introduced Act2Answer, a protocol that adapts VLM knowledge benchmarks to VLA evaluation, finding that VLA models perform well on simple concepts but struggle with richer semantic categories^[1]. Another study found that VLA models have multi-billion parameter architectures, imposing computational burdens during fine-tuning and real-time inference^[2]. To address this, a structural compression pipeline was introduced, reducing model depth by up to 50% and yielding a 40-50% reduction in training time and up to 30% faster real-time inference^[2]. However, VLA models still struggle with personalized commands, such as "bring my cup," where the robot must act on one specific instance among visually similar objects^[3]. To address this, researchers introduced VAP, a training-free perceptual adapter that equips frozen VLAs with top-down selective attention, consistently outperforming generic policies and token-learning baselines in success rate and correct-object manipulation^[3].