Differentiable Integrated Motion Prediction and Planning With Learnable Cost Function for Autonomous Driving.

Area of Science:

• Robotics and Artificial Intelligence
• Autonomous Driving Systems
• Machine Learning for Motion Planning

Background:

• Current autonomous driving systems often separate prediction and planning modules.
• Tuning the cost function for motion planning in autonomous vehicles is challenging.
• Predicting surrounding traffic participant behavior is critical for safe autonomous navigation.

Purpose of the Study:

• To propose a differentiable integrated prediction and planning (DIPP) framework for autonomous vehicles.
• To enable joint learning of prediction, planning, and cost function weights from data.
• To improve the safety, smoothness, and social compliance of autonomous vehicle trajectories.

Main Methods:

• Developed a differentiable nonlinear optimizer for motion planning within the DIPP framework.
• Integrated a neural network for predicting surrounding agent trajectories.
• Trained the framework on a large-scale real-world driving dataset to imitate human driving.

Main Results:

• The DIPP framework demonstrated superior performance in open-loop testing, producing human-like trajectories.
• Closed-loop testing showed the framework's effectiveness in complex urban scenarios and robustness to distributional shifts.
• Joint training of prediction and planning modules significantly outperformed separate training approaches.

Conclusions:

• The proposed DIPP framework offers a unified approach to prediction and planning for autonomous vehicles.
• Jointly learning prediction and planning enhances trajectory optimization and overall system performance.
• The learnable components are crucial for ensuring planning stability and achieving high performance in autonomous driving.