learning_material.md

note: 某些链接是非空开的

PNC算法学习

• Data-Driven/End2End

  • Notion: 数据驱动PNC && DayStore: Data-Driven Planning

  • Algorithm｜dd | VectorNet code

  • Algorithm｜e2e | Can Autonomous Vehicles Identify, Recover From, and Adapt to Distribution Shifts?, 2000

  • Algorithm｜e2e | Causal Confusion in Imitation Learning, 1900

  • Algorithm｜e2e | DTPP: Differentiable Joint Conditional Prediction and Cost Evaluation for Tree Policy Planning in Autonomous Driving, 2300

  • Algorithm｜e2e | VADv2: End-to-End Vectorized Autonomous Driving via Probabilistic Planning, 2402

  • Algorithm｜e2e | DriveVLM: The Convergence of Autonomous Driving and Large Vision-Language Models

  • Algorithm｜e2e | PiP: Planning-informed Trajectory Prediction for Autonomous Driving, 2000

  • Algorithm｜e2e | Scene Transformer: A unified architecture for predicting multiple agent trajectories,2103

    • 模型输入为perception object
    • 使用mask来支持同步做planning&prediction

  • Algorithm｜e2e | LOKI: Long Term and Key Intentions for Trajectory Prediction

  • Algorithm｜e2e | MP3: A unified model to map, perceive, predict and plan

  • Algorithm｜e2e | TNT

  • Algorithm｜e2e | DenseTNT

  • Algorithm｜e2e | MultiPath

  • Algorithm｜e2e | MultiPath++

    贡献有三点

    1. 精心设计下面几个方面提高效果：输入的表征及编码，融合编码及输出的分布。considering choices for input representation and encoding, fusing encodings, and representing the output distribution.
    2. 证明了以下几个方面对于行为预测很重要：稀疏编码，高效融合方法，基于控制的方法以及可学习的锚 （sparse encoding, efficient fusion methods, control-based methods, and learned anchors）
    3. we provided a practical guide for various tricks used for training and inference to improve robustness, increase diversity, handle missing data, and ensure fast convergence during training.

  • Algorithm｜e2e | Hydra-MDP: End-to-end Multimodal Planning with Multi-target Hydra-Distillation, 2406

• Algorithm｜MPDM

• Algorithm｜EUDM

• Algorithm｜Spatial-temporal Semantic Corridor(SSC)

• Algorithm｜EPSILON

• Algorithm｜MDP && Algorithm｜POMDP

• Algorithm｜Planning on a (Risk) Budget: Safe Non-Conservative Planning in Probabilistic Dynamic Environments

• Algorithm | Synthesis and Stabilization of Complex Behaviors through Online Trajectory Optimization(iLQR/DDP)

• Algorithm｜MultiPath && Algorithm｜MultiPath++

• Algorithm｜IDM Driver & MOBIL Model

• Algorithm｜MARC: Multipolicy and Risk-aware Contingency Planning for Autonomous Driving, 2308

• Algorithm | [Contingency Plan] Contingency Model Predictive Control for Automated Vehicles, 1907

• Algorithm | [Contingency Plan] Contingency Model Predictive Control for Linear Time-Varying Systems, 2102

• Algorithm | 使用branch MPC进行交互多模态运动规划 Interactive multi-modal motion planning with Branch Model Predictive Control, 2110

• Algorithm | 抵达集/可达集(Reach Set & Reachable Set) Bridging the Gap Between Safety and Real-Time Performance in Receding-Horizon Trajectory Design for Mobile Robots, 1809

• Algorithm｜Spatio-temporal Motion Planning for Autonomous Vehicles with Trapezoidal Prism Corridors and Be ́zier Curves(todo)

基础算法学习

• Algorithm｜图网络(GNN)

• Algorithm｜强化学习

• Algorithm｜状态栅格 state lattice planner

• Algorithm｜PWJ: Piecewise Jerk Optimizer(path & speed)

• Algorithm｜规划｜常用算法

  • A*
  • RRT&RRT*
  • 人工势场法(Artificial Potential Field)
  • Optimal Trajectory In Frenet Frame

• Curve

  • Bezier Curve 贝塞尔曲线
    1. 端点插值
    2. 凸包性：贝塞尔曲线
    3. hodograph性质: 它的导数依旧是贝塞尔曲线
    4. 固定时间间隔：每一段定义在[0, 1]间
  • Clothoid/Spiral（螺线）
  • Quintic Polynomial

• Spline 样条曲线

  • Cubic Spline
  • Quintic Spline
  • BSpline (Basis-Spline)
  • G2-Quintic Spline, code

• Control

  • LQR | 线性二次调节器: 状态转移关系是线性的，目标函数是二次的调节器.

    1. 设计线性的转移方程
    2. 调节矩阵 Q, R的值使得效果最优. Q 对角阵每个值是其对应状态维度的error惩罚weight. R 是 控制维度 weight
    3. 寻找最优的gain set: k = lqr(Q, B, Q, R)

  • MPC | 模型预测控制

    1. 估计/测量系统当前状态
    2. 基于转移方程和目标函数优化 控制序列 u(k), u(k+1), ..., u(k+N-1)
    3. 取最开始的u(k)控制输入到控制器 (滚动优化)

    MPC每一轮滚动优化可以使用LQR

  • iLQR(DeepNote - iLQG)/DDP Gitee-ddp

在线学习课程

• 课程｜多伦多大学-无人驾驶课程
• 课程｜深蓝学院-运动规划课程
• 课程｜深蓝学院-决策预测课程

编程

• C++/Python/Matlab
• Tensorflow/PyTorch