Learning-Based Behavior Planning for Automated Driving: Real-World Integration and Deployment

Noosaga Placements

• Learning-Based Motion Planningframework95%
  The paper proposes a learning-based motion planning approach, specifically a deep neural network for behavior planning, which falls under this framework.
  we developed a deep neural network to interpret complex traffic scenes and propose driving behaviorThe presented behavior planner is implemented as a deep neural network (DNN) in PyTorch
• Robot Motion Planningsubfield95%
  The paper focuses on behavior and trajectory planning for automated vehicles, proposing a hybrid architecture that combines learning-based and optimization-based planning methods.
  we propose a hybrid planning architecture that combines deep learning-based behavior planning with an optimization-based trajectory supervisionthe trajectory supervision module transforms the reference trajectory into a drivable and collision-free trajectory
• Safety-Critical and Verified Controlframework90%
  The trajectory supervision layer enforces safety constraints and ensures drivability, which aligns with safety-critical and verified control.
  an optimization-based supervision layer validates this proposal and enforces explicit drivability and safety constraintsThe trajectory supervision module safeguards the learning-based reference trajectory ... enforcing explicit safety, traffic-rule compliance, drivability, and vehicle-dynamic constraints
• Deep Learningsubfield90%
  The behavior planner is implemented as a deep neural network with transformer-based architecture, trained via imitation learning, and includes auxiliary tasks.
  The presented behavior planner is implemented as a deep neural network (DNN) in PyTorchadopts design principles from recent imitation-based transformer planners
• Optimal Control and Model Predictive Controlframework90%
  The supervision layer formulates and solves an optimal control problem (OCP) for trajectory optimization, which is a model predictive control approach.
  we developed a trajectory optimization framework that formulates and solves an optimal control problem (OCP)The optimization framework is built on acados, which is used to formulate the nonlinear OCP and generate solver libraries for online execution
• Robot Controlsubfield80%
  The architecture includes a trajectory controller and a supervision layer that enforces drivability and safety constraints, which are control-related components.
  we developed a cascaded proportional-integral-derivative (PID)-based Ackermann trajectory controllerthe trajectory supervision module ... enforces explicit safety, traffic-rule compliance, drivability, and vehicle-dynamic constraints
• Attention Mechanisms and Transformersframework85%
  The behavior planner uses attention-based scene fusion mechanisms and transformer backbones, which are part of attention mechanisms and transformers.
  adopts design principles from recent imitation-based transformer plannersincluding attention-based scene fusion mechanisms
• Ai Safetysubfield70%
  The paper addresses safety assurance, explainability, and trustworthiness of learning-based systems, and proposes a supervision layer to enforce safety constraints.
  their complex nature and lack of transparency can hinder explainability and trustworthiness and complicate safety assurancean optimization-based supervision layer validates this proposal and enforces explicit drivability and safety constraints
• Supervised Deep Learningframework85%
  The behavior planner is trained via supervised learning (imitation learning) on expert trajectories, which falls under supervised deep learning.
  predicts a reference trajectory based on a policy learned through supervised learning from selected datathe planner is trained by imitation: for each recorded scene, the network is trained to predict the corresponding expert trajectory
• Machine Learningsubfield60%
  The behavior planner is trained via supervised learning (imitation learning) on real-world driving data, which is a machine learning approach.
  predicts a reference trajectory based on a policy learned through supervised learning from selected datathe planner is trained by imitation: for each recorded scene, the network is trained to predict the corresponding expert trajectory
• Hybrid Deliberative/Reactive Architecturesframework80%
  The paper proposes a hybrid architecture that combines learning-based and classical planning, which is situated within learning-based robotics.
  we propose a hybrid planning architecture that combines the advantages of machine learning with the verifiability and the determinism of classical approachescombines deep learning-based behavior planning with an optimization-based trajectory supervision
• Learning-Based Controlframework70%
  The trajectory controller uses PID control with feed-forward terms, which is a learning-based control approach in the sense of model-based control.
  we developed a cascaded proportional-integral-derivative (PID)-based Ackermann trajectory controller

Abstract

Recent research in machine and deep learning has shown the potential of learningbased motion planning approaches to improve the driving behavior of automated vehicles, especially in complex environments. However, their complex nature and lack of transparency can hinder explainability and trustworthiness and complicate safety assurance. Motivated by these challenges, we propose a hybrid planning architecture that combines the advantages of machine learning with the verifiability and the determinism of classical approaches. Specifically, we developed a deep neural network to interpret complex traffic scenes and propose driving behavior, while an optimization-based supervision layer validates this proposal and enforces explicit drivability and safety constraints. We evaluate the learned planner's driving behavior in open-loop studies on real-world urban data, discuss system integration aspects for stable closed-loop operation, and report results from real-world deployment on our research vehicle karl..

Paper Context