Paper proposes Artificial Foveal Perception (AFP) to mitigate shortcut learning in robot foundation models

A paper by Sun et al. proposes Artificial Foveal Perception (AFP), a policy-agnostic lightweight module that constrains the visual attention of robot foundation models during fine-tuning to suppress shortcut learning. The authors argue that existing Vision-Language-Action (VLA) and World Action Model (WAM) pipelines still require task-specific fine-tuning in new robot scenarios, and models tend to rely on non-causal correlated features such as lighting and background texture, failing under distribution shift.