roboticscritiquebearishExisting visual-language-action models and world models struggle with accurate 3D geometry and physically meaningful forecasting for dynamic manipulationComputation and Language27 Jul 2026http://arxiv.org/abs/2607.01938v1