FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies
ResearchAI 88
Source: arXiv cs.AIPublish time unverified
arXiv:2605.27284v3 Announce Type: replace-cross Abstract: Vision-Language-Action (VLA) models are increasingly expected to not only complete robot tasks, but also follow human instructions about how those tasks should be executed. However, existing robot datasets usually pair trajectories with coarse goal-level language, leaving execution-critical details such as active arm, approach direction, and contact region unspecified.