Qwen-Drive’s Trajectory Pipeline and Integration Boundaries
A Juejin analysis of Qwen-Drive’s public code and technical report maps how its Qwen3.5-4B visual-language backbone feeds separate BEV perception and planning components. The planning interface can return multiple candidate trajectories, while the singular trajectory property simply selects the first candidate rather than scoring it; the reported default is 50 future points per trajectory over five seconds at a 10 Hz data interval. The review also details camera, motion-state and navigation inputs, training stages, and the boundary between trajectory generation, candidate selection, and downstream control.
The source identifies concrete integration requirements for engineers using Qwen-Drive, including preserving the full candidate array, interpreting coordinates and time correctly, and providing the required multiview and vehicle-state inputs.
Qwen -Drive 到底是什么:4B 级 VLM 如何成为驾驶策略底座
如果你准备把 Qwen-Drive 接入一个驾驶研究服务,最先要决定的接口可能很小:服务返回一条轨迹,还是一组候选?假设同事顺手用了 result.trajectory,界面画出一条五秒路线,接入就算完成了吗?
公开源码里,这个单数属性只有一行:return self.trajectories[0]。它取出第一条采样结果,没有在这里评价其余候选,更没有把“第一条”认证成“最合适的一条”。这处细节,比“4B 模型 会开车”的描述更能帮助工程师判断它承担了什么工作:Qwen-Drive 的规划路径生成未来自车轨迹;候选怎样进入下游,仍然需要一份明确的接入安排。1
沿这个返回值向前看,可以看清共享 视觉语言 主干、驾驶训练、连续轨迹专家和输入协议各自的作用。本文回答一个问题:接入 Qwen-Drive 时,应该把它当作什么组件,在哪个输出位置划清边界?最后形成的是一份待执行的输入—轨迹—下游接入记录。本文核对固定源码和官方论文,没有加载模型、取得 GPU 性能或运行车辆与仿真。