Mobileye's CES 2025 presentation outlines their strategy for revolutionizing transportation through autonomous driving. They discuss three approaches to achieving Level 5 autonomy: prioritizing precision then recall (like Waymo), prioritizing recall then precision (like Tesla), and Mobileye's own approach of achieving high precision first (with their "chauffeur" system) before increasing recall. A key element of Mobileye's strategy is their focus on redundancy and safety, detailed in a recently published safety report. The presentation also covers their technological advancements, including their new iq6 chip and Imaging radar, and their commitment to cost-effective mass production.
- Mobileye's Five Pillars to Revolutionize Transportation
- Safety: The system must go beyond MTBF and focus on eliminating "unreasonable risks."
- They define four types of potential errors: planning errors ("lapses of judgment"), interpreation errors, identifiable errors (hardware or software failures), and "black swan" errors (unpredictable and extremely rare failures).
- They use a "primary-guardian-fallback" methodology for decision making that requires multiple subsystems to fail before a significant failure occurs. This boosts precision through design, not data. This is an RSS based architecture.
- Technology Stack: The complete stack must include software, silicon, and sensors, not just software.
- Scalability: Achieving out-of-the-box performance across different geographies is vital.
- Productization / Execution: Moving from demos to mass-produced, robust products is a significant challenge.
- Cost: Developing a system that is both autonomous and affordable for consumers is a must.
- Redundancy is Key:
- Mobileye emphasizes creating "redundancies" within its system to mitigate failures and increase precision. Redundancy is used to address both identifiable errors and black swan errors.
- This includes redundant hardware (multiple EyeQ chips), software, and sensors (cameras, radars, lidars).
- Their approach emphasizes sensor modality redundancies to ensure that failure involves at least two subsystems before any negative consequences.
- Mobileye's Technological Approach:
- EyeQ™6H Chip: A new generation chip with a 10x increase in frames per second compared to its predecessor (EyeQ™5H), leading to better algorithm performance, higher compute density, and a projected MTBF improvement. Expected MTBF of 500-1000 hours from the camera system alone.
- Imaging Radar: A key component in their eyes-off solution, with significantly higher performance than competitors in terms of dynamic range and object separation. It's a "software defined" radar that allows for much improved resolution, giving it high accuracy and enabling virtual channels.
- REM (Road Experience Management): Their crowdsourced mapping technology that is crucial for out-of-the-box performance anywhere in the world. They have collected 56.6 billion miles of data through REM by 2024, with 29.6 billion miles collected in 2024 alone.
- RoadX: A transformer network that uses drone view images, REM, and SD maps to output drivable paths, boundaries, and semantics.
- Vidar (Visual Lidar): A method of creating a 3D depth perception from camera data using triangulation.
- Compound AI System: Using different AI models and techniques (end-to-end, decomposable, injecting abstractions) and fusing them together in a unified and redundant system
