A new research paper, "VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding," describes a method to accelerate visual planning in AI systems used for robotics and autonomous driving. The paper, published on arXiv, details how VFlash employs a technique called Flash Primitive Decoding to generate visual plans more efficiently than existing autoregressive vision-language models.Traditional autoregressive models generate visual plans, such as points, bounding boxes, and trajectories, one coordinate token at a time. This sequential generation can become computationally expensive, particularly for more complex or detailed predictions. VFlash addresses this by exploiting the inherent structure of visual primitives. Instead of generating individual tokens, VFlash predicts the coordinates of an entire primitive simultaneously using a lightweight "drafter."The drafter in VFlash is designed to generate a complete visual primitive in a single forward pass. Its layout adapts to the specific type, dimensionality, and length of the primitive, allowing it to handle diverse outputs like compact bounding boxes and extended, variable-length trajectories. This parallel processing contrasts with the step-by-step approach of autoregressive models.A key aspect of VFlash is that it does not require reproducing an autoregressive model's exact token sequence to achieve spatial accuracy. The system trains on annotated primitives and accepts predictions that are structurally valid, even if the token sequence differs from an autoregressive target.Across various evaluated 2D and 3D robot trajectory tasks, VFlash consistently reduced prediction errors compared to autoregressive decoding methods. Preliminary measurements for 3D point motion forecasting indicate a generation speed increase of approximately 25 times.Furthermore, a world action model that utilized VFlash plans achieved a 70.2% average success rate across four MolmoSpaces task suites. This compares to a 62.6% success rate for a baseline model. The researchers also noted that VFlash reduced inference latency by about 80% on NVIDIA's Jetson Thor and DGX Spark platforms in a downstream robot manipulation pipeline.VFlash is described as a generalist decoding framework for physical AI, capable of handling both text, such as sub-task plans, and various visual primitives including bounding boxes, variable-length trajectories, and dense 3D motion predictions. The framework's ability to align prediction blocks with the known structure of visual primitives enables parallel decoding of common structured outputs.