论文提出 BIPP，把自动驾驶中的预测与规划从单向串行流程改为并行、迭代的双向交互流程。各阶段共享未来上下文，联合利用历史运动和更新后的意图，并采用规则生成横向轨迹、学习策略控制纵向动作的混合规划方案。

Safe and e fficient autonomous driving in dense urban environments requires both accurate modeling of multiagent interaction and robust closed-loop execution. However, many existing methods still use a one-pass sequential predictthen-plan implementation, in which the current plan does not feed back to the current surrounding-agent prediction and the fixed update order introduces a structural asymmetry between the autonomous vehicle (A V) and surrounding agents. In addition, purely learning-based planners often su ffer from distribution shift and error accumulation during closed-loop execution, and consequently rely on heavy post-processing to maintain safety. To address these limitations, we propose BIPP, a Bidirectional Interactive Prediction and Planning framework for autonomous driving. BIPP reformulates conventional sequential decision making as a parallel-iterative process, in which prediction and planning are performed in parallel within each stage and progressively refined across stages using shared future context. A unified bidirectional interaction modeling mechanism jointly leverages historical motion evidence and iteratively updated future intentions to capture mutual adaptation between the A V and surrounding agents. Moreover, a hybrid planning strategy combines rule-based lateral trajectory generation with a learningbased longitudinal policy to improve feasibility, adaptability, and closed-loop performance. Closed-loop experiments on nuPlan and interPlan show that BIPP achieves the best overall performance among the compared baselines in the evaluated benchmark settings, with notable gains in highly interactive and long-tail scenarios.

传统先预测再规划的单向流程中，当前规划不会反馈给周边车辆预测，固定顺序也造成自车与周车建模不对称；纯学习式规划还容易受分布偏移和闭环误差累积影响。

To address these limitations, we propose BIPP, a Bidirectional Interactive Prediction and Planning framework for autonomous driving.

在同一迭代框架中让预测结果与规划意图相互更新，并以混合规划兼顾可行性、适应性和闭环执行安全。

To address these limitations, we propose BIPP, a Bidirectional Interactive Prediction and Planning framework for autonomous driving.

每一阶段并行执行预测和规划，再用共享的未来上下文逐步细化；通过双向交互模块联合历史轨迹与更新后的未来意图，并结合规则式横向轨迹和学习式纵向策略。

To address these limitations, we propose BIPP, a Bidirectional Interactive Prediction and Planning framework for autonomous driving.

nuPlan 与 interPlan 闭环实验中，BIPP 在论文评测设置下总体优于对比基线，收益集中在交互密集和长尾场景；摘要未给出统一提升幅度。

Closed-loop experiments on nuPlan and interPlan show that BIPP achieves the best overall performance among the compared baselines in the evaluated benchmark settings, with notable gains in highly interactive and long-tail scenarios.