VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding
A world action model using VFlash plans reached 70.2% average task success, versus 62.6% for the baseline.
VFlash predicts the coordinates of a visual primitive together, rather than generating them one token at a time. It accepts structurally valid predictions without requiring an exact match to an autoregressive model’s token sequence. The method reduced prediction errors across the evaluated 2D and 3D robot trajectory tasks. Preliminary single-rollout measurements showed approximately 25× faster generation for 3D point motion forecasting.
HF Daily Papers' note
VFlash predicts the coordinates of a visual primitive together, rather than generating them one token at a time. It accepts structurally valid predictions without requiring an exact match to an autoregressive model’s token sequence. The method reduced prediction errors across the evaluated 2D and 3D robot trajectory tasks. Preliminary single-rollout measurements showed approximately 25× faster generation for 3D point motion forecasting.
HF Daily Papers' note
score 5