Vision-language-action (VLA) models combine the world knowledge of pretrained vision-language models with robotic control, enabling robots to follow natural-language instructions. However, a key challenge is the structural mismatch between the semantic representations learned by VLMs and the continuous, low-level actions a robot must generate. This gap can limit how effectively semantic understanding translates into physical manipulation.
The paper introduces AffordanceVLA, a model designed to overcome this mismatch by incorporating affordance-aware understanding. Affordances—the possible actions an object offers—provide a bridge between high-level semantics and concrete motor commands. By making the model explicitly aware of what actions are feasible in a scene, AffordanceVLA aims to empower more accurate and contextually appropriate action generation.
According to the abstract, the model leverages the rich world knowledge of pretrained VLMs while addressing the structural mismatch that has hindered prior VLA approaches. The paper is available on arXiv and represents a step toward more robust instruction-following robots that can better align what they see and understand with what they actually do.