E2E + Rule-Based Hybrid Autonomous Driving Platform on Autoware

[YSH-research]

Vision

Implement GenAD-based E2E + Rule-Based hybrid driving on Autoware,
and demonstrate it in both CARLA simulation and real-vehicle driving.

Background & Motivation

Autoware's Autonomy 2.0 direction aims to integrate learning-based
end-to-end driving models. However, pure E2E models alone cannot
reliably handle all edge cases. A hybrid approach combining E2E
planning with a rule-based safety layer can achieve both performance
and reliability for real-world deployment.

I am working on a real-vehicle E2E driving project, and I am
leading the Autoware-based E2E integration effort, having prior
experience deploying Autoware on real vehicles. I previously worked
on integrating E2E models into Autoware through the
autoware_tensorrt_vad project
at a basic implementation level, and now aim to pursue this
systematically as a full Autoware Project.

Components

Component	Description	Status
autoware_tensorrt_genad	GenAD TensorRT-accelerated package	#12317
hlc_publisher	High-Level Command publisher for E2E models	Planned
e2e_trajectory_selector	E2E + Rule-Based trajectory switching selector	Planned
e2e_evaluation_tools	ros2top-based resource analyzer and E2E evaluation tool	Planned

Roadmap

Phase 1 — Package Development

• Train and validate GenAD on B2D dataset (GenAD_no_hlc)
• GenAD TensorRT quantization (FP16) using DL4AGX — nearly complete
• Create autoware_tensorrt_genad package (#12317)
• Create hlc_publisher package
• Create e2e_trajectory_selector package
• Create e2e_evaluation_tools (ros2top-based)

Phase 2 — Validation & Driving

• CARLA simulation validation
• GenAD-based E2E + Rule-Based hybrid pipeline integration
• Real-vehicle driving demonstration

[Max-Bin]

@YSH-research
Thank you for your excellent plan! We look forward to hearing about your progress.