TJ Autonomous Driving Project

Introduction

During my time at Tongji University, I took part in an innovative project bridging the virtual and physical worlds for autonomous vehicle (AV) safety testing. Our team developed a novel device designed to evaluate the performance of commercial self-driving systems under hazardous conditions that are difficult, dangerous, or ethically challenging to recreate on real roads. Of which we developed is a specialized device that physically obscures vehicle’s own cameras. Crucially, this device doesn’t leave the car “blind.” Instead, it integrates high-resolution and high-FPS displays that dynamically display meticulously simulated virtual driving environments. This puts the actual car in a virtual environment that allows people to subject its real sensor processing and decision-making algorithms to entirely virtual, yet highly realistic, hazardous situations within a safe laboratory or controlled track setting.

The primary advantage of this system is its ability to safely generate and test edge cases and corner cases:

• Safety: Eliminates the risk inherent in testing dangerous scenarios (e.g., high-speed collisions, pedestrian jaywalking in low visibility) with real vehicles or actors.
• Repeatability & Control: Scenarios can be reproduced with pixel-perfect precision, enabling reliable A/B testing and regression analysis.
• Scalability: Allows for the rapid testing of thousands of scenario variations far beyond the scope of physical test tracks.
• Complexity: AI can generate novel, complex scenario combinations that might be statistically rare but critically important for safety validation.

---

Within the software team, my specific responsibility focused on establishing the foundational scenarios for testing AEB systems. My task was to accurately reproduce the standardized test scenarios defined in the C-NCAP protocol within the virtual environment via CARLA and Scenario Generator Python library to create a .xosc file for each test scenario.

My Workflow:

1. Protocol Analysis: Deeply analyzing the C-NCAP AEB test specifications
2. CARLA Environment Setup: Configuring the virtual world in CARLA to match the test track layouts and environmental conditions specified by C-NCAP.
3. Scenario Scripting: Using Python and the scenario generator libraries to define the precise actor behaviors , trigger conditions, and timing required by each C-NCAP test case.
4. Export: Generating the final, standardized OpenSCENARIO (.xosc) files representing each faithfully recreated C-NCAP AEB test scenarios. These files became the core input for our physical-virtual test device’s simulation engine.

Impact and Next Steps

The scenarios I reproduced served as the critical baseline for the project. Once validated within the virtual CARLA environment and exported as .xosc, these scenarios were then fed into the larger system. The project’s AI-driven component could then take these fundamental building blocks, parameterize them, and perform intelligent permutations and combinations to generate vastly more complex and diverse hazardous situations. These novel scenarios were ultimately displayed on the device’s screens, testing the real vehicle’s AEB system against challenges far exceeding the original standardized tests.