Utopi
a

Integrated Decision and Control: Toward Interpretable and Computationally Efficient Driving Intelligence

Abstract

  • Backgroud

    Current mainstream methods suffers high time complexity or poor interpretability and adaptability on real-world autonomous driving tasks.

  • Our contribution

    integrated decision and control (IDC) :a interpretable and computationally efficient framework

  • How:

    decompose into 2 parts: [ SPP: static path planning ] + [ DOT: dynamic optimal tracking ]

    • SPP only considers static traffic elements -> generates candidate paths (each formulated with a constrained optimal control problem (OCP)),

    • DOT track the optimal path while considering the dynamic obstacles (each candidate path is optimize separately )

    • model-based RL algorithm served as an approximate-constrained OCP solver to ease computation burden

      • 具体来说，将所有路径的 OCP 一起考虑以构建一个完整的 RL 问题，然后以价值网络和策略网络的形式离线求解，分别用于实时在线路径选择和跟踪。

  • result

    结果表明，与基线方法相比，IDC 具有一个数量级的在线计算效率，以及更好的驾驶性能，包括交通效率和安全性。此外，它在不同的驾驶场景和任务之间产生了很好的可解释性和适应性。

Introduction

研究技术回顾

two technical routes for the decision and control of automated vehicles:

  1. decomposed scheme

    splits into several serial submodules:

    • scene understanding
    • prediction
      • behavior selection (finite-state machine, decision tree ),
      • trajectory planning
        • optimization-based -> high computational complexity
        • search-based -> low-resolution & barely consider dynamic obstacles.
      • Xin's Enable faster and smoother spatio-temporal trajectory planning for autonomous vehicles in constrained dynamic environ- ment combined them together -> requires a large amount of human design & cannot guarantee real-time performance
    • control

  2. end-to-end scheme.

    直接从感知模块给出的输入计算预期指令，不依赖于标记数据。

    有限的安全性能和较差的学习效率，主要用于特定任务，其中需要一组复杂的奖励函数来为策略优化提供指导，例如前往目的地的距离、与其他道路使用者或场景对象的碰撞以及维护舒适性和稳定性，同时避免极端加速、制动或转向。

This article：

proposes an integrated decision and control (IDC) framework for automated vehicles, which has great interpretability and online computing efficiency and is applicable in different driving scenarios and tasks

3 parts of contributions:

  1. ​ IDC framework:
    • 高级静态路径规划：仅考虑静态约束（例如道路拓扑和交通信号灯）生成多条路径。
    • 低级动态最优跟踪：选择最优路径并考虑动态障碍物对其进行跟踪，其中针对每个候选路径构建和优化有限视域约束最优控制问题（OCP）
    • 优势：
      • 高效计算：because it unloads the heavy online optimizations of the constrained OCPs in offline using RL
      • 可解释：because the value and policy NNs are the approximation for the optimal cost and action of the constrained OCP
      • One more thing: the formulated OCP in the IDC is task independent with tracking objective and safety constraints, making it applicable among a variety of scenarios and tasks.
  2. model-based RL algorithm: generalized exterior point method (GEP)
    • 它首先构造一个涉及所有候选路径的广泛问题，并将其转换为一个不受约束的问题，并对安全违规进行惩罚。然后，通过交替执行梯度下降和扩大惩罚来获得近似可行的最优控制策略。
    • GEP 可以明确地处理具有大规模状态约束的问题，并在模型的指导下有效地更新 NN。
  3. verify in simulations and real-world road: online computing efficiency, safety, task adaptation

Intergrated and Decision and Control Framwork

  • upper layer only consider the static problems: traffic light etc & not include time information, generating serveral candidate path
  • considers the candidate paths and the dynamic information, a constrained OCP is formulated and optimized for each candidate path to choose the optimal path and find the control command. 我们方法的核心是用 RL 离线训练的两个 NN 的前馈来代替所有昂贵的在线优化。

terminology

什么是离线优化、在线优化、增量优化？

离线常用优化器 https://zhuanlan.zhihu.com/p/360258616

Interior-point method

https://zhuanlan.zhihu.com/p/34426575

处理带约束优化问题