Camera-only crusade: autonomous driving goes all-in on vision, slashing costs and redefining safety

In the early days of autonomous trucking, particularly around 2016 to 2018, the technology stack was heavily lidar-centric, with companies striving to develop long-range lidar systems capable of detecting objects 300 to 400 meters away to accommodate the extensive stopping distances of heavy trucks. Cameras played a secondary role due to their limited vision capabilities at the time, and the AI driving these systems relied on handcrafted algorithms rather than the advanced machine learning models seen today. Despite the rapid evolution in AI and sensor fusion techniques over the past decade, foundational safety engineering principles and hardware robustness established during this initial phase have remained remarkably consistent, underscoring the enduring importance of rigorous validation and durable sensor platforms in autonomous trucking.

By mid-2026, a clear industry pivot toward vision-first autonomous driving technologies has emerged, led by companies like Autobrains, Humble Robotics, and Mars Auto. Autobrains is championing vision-only systems aimed at mass-market vehicles, deliberately eschewing lidar to achieve human-level driving performance affordably, with plans for mass-scale deployment within five years. This democratization contrasts with earlier premium-focused approaches, signaling a strategic shift to make autonomy accessible to everyday drivers.

Humble Robotics exemplifies the rapid maturation of camera-based AI vision models, leveraging pre-trained networks trained on vast internet data to drastically reduce engineering time and enhance scenario recognition—from traffic cones to stop signs—without relying heavily on lidar. While their current vehicles employ a multisensor setup including lidar and radar for safety, the company is increasingly bullish on transitioning to camera-only systems, echoing long-standing industry advocates like Elon Musk who have predicted this technological tipping point for over a decade.

Mars Auto is pushing the boundaries of camera-only autonomy in the heavy trucking sector, deploying a single neural network that processes raw video to perform perception, judgment, and control without HD maps or lidar. This approach slashes sensor costs dramatically—from rivals’ $229,000 to about $6,500 per vehicle—while proving its mettle through a record-setting 3,379 km autonomous freight haul across the U.S. With plans to commercialize driverless trucks by 2028 and expand paid autonomous trailer operations in Korea by late 2024, Mars Auto is also developing MarsNet 3, a next-gen Level 4 system extending camera-only autonomy into urban environments without reliance on HD maps.

By mid-2026, Valeo revolutionized radar sensing with a breakthrough platform delivering 50 times the resolution of previous generations, enabling radar to generate detailed environmental maps once exclusive to lidar systems. This leap in sensor fidelity, combined with deep learning-based sensor fusion, forms a 360-degree safety shield around vehicles, allowing millimeter-accurate maneuvers critical for autonomous parking and advanced driver assistance, thereby significantly enhancing safety and precision in vision-first autonomous architectures.

Valeo’s radar systems embody a forward-thinking modular hardware design paired with over-the-air update capabilities, allowing vehicles initially equipped with Level 2+ autonomy to seamlessly upgrade to Level 3 functionalities as regulatory landscapes and software validations mature. This strategic flexibility not only future-proofs investments for manufacturers and consumers but also accelerates the deployment of advanced autonomous features in real-world conditions.

Through its 'Elevate 2028' initiative, Valeo is transitioning from a component supplier to a pivotal system integrator, actively shaping centralized AI software architectures that harmonize hardware excellence with sophisticated software intelligence. This evolution underscores Valeo’s commitment to driving the next wave of autonomous driving innovation by delivering holistic solutions that integrate high-resolution sensors with advanced AI fusion, setting new industry standards for safety and operational efficiency.

StradVision exemplifies a pragmatic, incremental approach to autonomous driving by targeting Level 2 to Level 2+ ADAS markets with a clear path toward profitability by 2028. Leveraging its SVNet object recognition software, the company has already equipped over 5 million vehicles globally across China, India, the US, and Germany. This growth is underpinned by a cost-effective scalability strategy that emphasizes compatibility with around 30 semiconductor platforms and an energy-efficient AI vision system, enabling broad adoption without the need for expensive hardware.

VinFast’s strategy mirrors this incremental ethos, focusing on stepwise enhancements from Level 2 to Level 2++ autonomy through rigorous real-world validation and software refinement. By prioritizing a camera-first perception system over costly LiDAR and radar arrays, VinFast aims to democratize autonomous technology with its Robo-Car platform, which integrates seven production-grade cameras and a high-performance computing unit capable of 20 trillion operations per second. This approach not only reduces costs but also reinforces driver responsibility, as Level 2 systems assist rather than replace human control.

Strategic partnerships are central to VinFast’s incremental deployment, notably its collaboration with AI innovator Autobrains to accelerate development of safer, more intelligent Level 2++ autonomous driving technologies. This alliance supports the evolution of a simplified autonomous vehicle architecture currently undergoing pilot testing on VF 8 and VF 9 models in controlled Hanoi zones, demonstrating VinFast’s commitment to pragmatic, scalable solutions validated in real-world environments before broader market rollout.