NVIDIA launches VFlash for faster robotic visual planning
DoNews reports that NVIDIA introduced VFlash, a parallel decoding framework designed for rapid and accurate visual planning in robotics and autonomous driving. Its Flash Primitive Decoding method generates complete trajectories and bounding boxes as independent blocks, with reported results including up to 12× faster robot action decoding and 5.1× faster end-to-end robot inference on an H100 at batch size 1. The framework runs through TensorRT-Edge-LLM on H100, DGX Spark, and Jetson Thor, though all cited results come from benchmarks rather than real-world deployments.
VFlash targets a key bottleneck in physical AI by speeding up geometric outputs used for robot actions and autonomous-driving planning, while its reported error reductions and edge-hardware support make it relevant to real-time deployment.
11月6日消息,英伟达于近日正式推出面向高速精准视觉规划的通用解码框架VFlash,通过革新视觉空间信息生成逻辑,为机器人、自动驾驶系统带来推理速度与任务性能双重提升。
传统大模型普遍采用自回归逐 Token 生成模式,该模式适配文本生成,但并不完全适配机器人场景。机器人执行抓取、移动等动作时,需要输出大量空间坐标、运动航点、三维轨迹等具备强几何属性的数据;逐一生成坐标不仅耗时漫长,还容易产生级联误差,制约机器人实时响应能力。
VFlash 核心创新为闪速原语解码(Flash Primitive Decoding),不再逐个输出坐标点位,而是把轨迹、包围框等完整几何对象视作独立解码块。系统先通过工具调用声明输出长度,明确视觉基元类型,在冻结视觉语言模型(VLM)上下文条件下,依靠轻量化草稿模型一次性并行预测完整空间信息,再交由目标 VLM 完成特征提取,供给下游动作模型使用。