Idea
Platform generating LiDAR-invisible 3D objects to test autonomous vehicle detection safety.
Research Paper
Core Innovation
This paper presents Phy3DAdvGen, the first method to generate physically realizable 3D adversarial objects that cause complete LiDAR invisibility. Unlike prior perturbation-based attacks, it optimizes text prompts and object combinations to produce realistic, deployable models that evade multiple state-of-the-art detectors in simulation and real-world tests.
Why It Matters
Autonomous vehicles rely on LiDAR for safe navigation, but undetected objects pose critical safety risks. This technology enables comprehensive testing by generating physically realizable adversarial 3D objects that evade detection, helping manufacturers identify and fix vulnerabilities before deployment, thus enhancing road safety and trust in autonomous systems.
Market Size (TAM)
$10–20B TAM for autonomous vehicle safety and testing tools; $2–5B SAM from AV manufacturers and sensor developers. Driven by increasing AV deployment and regulatory safety requirements.
Potential Customers & Pain Points
- Autonomous vehicle manufacturers–Need robust LiDAR detection testing
- LiDAR sensor developers–Require vulnerability assessment tools
- Autonomous driving software companies–Need to improve detection reliability
- Safety regulators–Require validation of detection systems
- Simulation platform providers–Need realistic adversarial scenarios.
Business Model
B2B SaaS platform offering simulation and physical adversarial object generation tools via subscription and licensing to AV manufacturers, sensor developers, and safety regulators.
Competitive Landscape
- Waymo Safety Testing
- NVIDIA DRIVE Sim
- Aeva LiDAR Testing
- Luminar Technologies
Implementation Challenges
- Physical realization complexity of adversarial objects
- Integration with diverse LiDAR systems
- Regulatory acceptance of adversarial testing methods
Validation Strategy
- Demonstrate evasion success across multiple commercial LiDAR detectors
- Pilot integration with leading autonomous vehicle manufacturers
- Conduct real-world physical environment tests
- Collaborate with regulatory bodies for safety validation standards
Research Paper Overview
OBJVanish: Physically Realizable Text-to-3D Adversarial Generation of LiDAR-Invisible Objects
Summary
This paper introduces Phy3DAdvGen, a novel text-to-3D adversarial generation method that creates physically realizable 3D pedestrian models invisible to LiDAR-based detectors. It systematically optimizes object topology, connectivity, and intensity within CARLA simulation and real environments, exposing vulnerabilities in state-of-the-art autonomous driving detection systems.