Tesla has an internal document called “Radar Saves Us.” On July 2, the National Highway Traffic Safety Administration (NHTSA) sent Tesla a letter demanding to know how that document was developed and why the company kept telling the public its camera-only system could handle any condition. The request ties directly to nine “Full Self-Driving” crashes, every one of them in conditions cameras are known to struggle with: sun glare, fog and airborne dust. It may be the clearest signal yet that a federal safety regulator has doubts about the vision-only approach.
To understand why that matters, it helps to take a step back and look at how self-driving vehicles are actually supposed to perceive the road.
Why Tesla Dropped Radar From Its Sensor Suite
Back in 2021, when Tesla announced that they were going vision-only and removing radar from their sensor suite, experts hesitated. They did not understand how that could possibly work since cameras notoriously do not function well in low visibility and challenging weather scenarios. How could a vehicle that is going to be driven by millions of people and be confronted with so many different complex scenarios on the road completely rely on just one sensor?
For years, our company has argued that a camera-only approach carries real safety risk, especially in reduced-visibility conditions like sun glare, fog or airborne dust. True safety comes from pairing vision-based sensors with complementary technologies like ultra-high-resolution radar. Different sensors capture different kinds of data. When one is limited, another is still there to keep detection accurate. This is standard practice across the industry. Every sensor has strengths and limitations. A well-designed sensor suite overlaps them. When one sensor falls short, another compensates.
Cameras, LiDAR and Radar: Strengths and Blind Spots
The camera is a foundational sensor. It reads the colors and two-dimensional shapes essential for lane markings and classification. But cameras struggle in low light. When an oncoming vehicle’s high beams saturate the image, they suffer from partial observability. A guardrail can hide a vehicle’s true shape, sometimes producing a false negative that the algorithm has to correct for.
LiDAR offers outstanding spatial resolution, differentiating pedestrians, bicycles and vehicles with remarkable accuracy in urban settings. But its performance drops sharply in rain, fog, snow, smoke or dust, since light waves scatter and get absorbed by water droplets or snowflakes, making it less reliable in regions with unpredictable weather.
That’s where high-resolution radar comes in.
How High-Resolution Radar Closes the Perception Gap
High-resolution radar breaks traditional radar limitations, offering ultra-high resolution that enables precise target differentiation. Even in dense urban traffic, inclement weather or poor lighting conditions, the radar can seamlessly classify and identify pedestrians, vehicles and other objects, consistently delivering reliable detection and tracking capabilities.
High-resolution radar compensates for the challenges that LiDAR and camera technology face. It can detect low-reflectivity objects, like pedestrians at night or road debris, effectively addressing the shortcomings of other sensors. This is the level of detail that makes eyes-off driving possible at highway speeds. Drivers can trust that lost cargo or debris in the road has already been detected and identified, not just assumed away.
When high-resolution radar is paired with cameras, it enables a comprehensive perception framework. Advanced high-resolution systems provide long-range, high-accuracy object detection, empowering drivers or autonomous systems to anticipate and respond to potential hazards more effectively. They can supply real-time data about object distance, speed and travel direction, all critical information for safe navigation and decision-making during high-speed maneuvers in any weather and lighting condition.
Why Sensor Resolution Matters for AI-Driven Perception
It is not only about having multiple types of sensors, either. The level of detail each one captures matters just as much, especially for artificial intelligence (AI). AI cannot manufacture information a sensor never measured. High-resolution radar directly gives perception algorithms richer, more densely sampled information about the physical world, such as distance, velocity and shape data captured. Lower-resolution systems can use AI to enhance resolution and infer missing detail, but inference is not measurement. The less information a sensor captures, the more the AI has to rely on learned assumptions to fill the gaps. On the road, an assumption can equate to risk.
Inside NHTSA’s Investigation Into Tesla’s “Radar Saves Us” Document
So did Tesla not know this? That is exactly what NHTSA is trying to determine: how and why Tesla decided to drop radar from its sensor suite, and whether the company accurately portrayed the system’s capabilities to the public.
In the letter NHTSA sent to Tesla, they requested any data, communications and documentation around how the “Radar Saves Us” document was developed. The NHTSA Office of Defects Investigation wrote in their letter:
This document appears particularly relevant to related design changes to the degradation detection system and Tesla’s understanding of the limitations imposed by a vision-only OEDR system and how they relate to the reduced visibility conditions that the Subject Vehicles encountered in each of the nine (9) countable crashes.
The letter also highlights the term “regressions” that Tesla used frequently as a metric in reports, referring to system performance and requests more details, including the definition of “regressions” and how they identified and measured them. NHTSA also asks Tesla to address previous marketing statements that were made, including a statement the company put out on X.com stating “FSD Supervised can drive in whiteout conditions” and a post by Elon Musk on X.com that states “Tesla FSD can operate in all conditions.” (This was in response to another post that shared a video stating that FSD struggles in difficult weather conditions.) These requests show that NHTSA is seriously investigating the safety of a vision-only approach.
According to reporter Fred Lambert from Electrek, who broke this story and has been following it closely, Musk has previously privately texted him that “vision with high-resolution radar would be better than pure vision,” indicating that he was aware that camera alone was not the safest solution.
The Case for Sensor Fusion Over Vision-Only Systems
At Arbe, we have never believed in automotive safety that relies on vision alone. Perception radar, paired with cameras, is what gives AI the redundancy and the richer, real-world data it needs to be safe today and capable tomorrow.
Musk’s argument in 2021 was that “humans drive with eyes and biological neural nets…so (it) makes sense that cameras and silicon neural nets are (the) only way to achieve a generalized solution to self-driving.” The problem with that mindset is that autonomous driving should not aim to be equal to human drivers. People will not forgive machines for collisions, the way they would forgive a human for making a human error. The automotive industry must aim for superhuman driving that will eventually eliminate automotive collisions completely. And for superhuman driving, you need superhuman sensors that can “see” beyond human capacity.
But don’t ask us. Ask NHTSA. They seem to have the same idea.