fexponential is an independent research initiative focused on safety and interpretability in Embodied AI.
We are currently developing an open-source framework for Mechanistic Interpretability in Vision-Language-Action (VLA) models. Our research aims to address critical reliability failures, such as temporal blindness in long-horizon tasks and the inability to adhere to dynamic safety constraints.