Tesla has extended the end of its Robotaxi service hours in Austin, Texas, from 10 p.m. to 11 p.m. In an October 3 post on X, Elon Musk said that avoiding pets that are hard to see at night is a major challenge the company is working on now. In the post, as reported by Electrek, he gave a specific example: trying to avoid "a gray kitten on a gray road in the dark."
The problem of spotting a small animal at night also ties into the broader debate over how autonomous vehicles should use cameras, LiDAR and radar. But safety does not depend only on whether sensors can detect an obstacle. It also matters whether the vehicle can recognize when it can no longer perceive its surroundings well enough, and then switch to a safe behavior such as slowing down or stopping.
Why is a kitten on a dark road so hard to see?
Musk's example combines two factors: very little ambient light, and an object that blends into its background.
An animal that would be easy to pick out in a bright setting, thanks to differences in color and outline against the road, offers far less information for a camera image to work with when a gray body lies on dark gray pavement. Darkness and low contrast with the background are separate conditions, and each makes recognition harder.
Tesla itself says visibility affects the performance of its consumer driver-assistance system, FSD (Supervised).
The Model 3 owner's manual states that visibility is critical for FSD (Supervised) to work properly, and that low light, rain, snow, direct sunlight, fog and similar conditions can significantly degrade performance.
This is not only a night-time issue. It applies to any situation in which the cameras cannot gather enough visual information from the surroundings.
However, that manual describes FSD for consumer vehicles, which assumes constant supervision by the driver. It is not a document measuring Robotaxi's obstacle-detection performance at night.
Musk's post likewise gives no figures, such as at what light level, from how many meters away, or for what size of animal detection is possible. Because he named small-animal detection in low light as a challenge, it cannot be concluded that the system is entirely unable to recognize small animals at night.
Nor does the extension of Robotaxi's end time from 10 p.m. to 11 p.m. in itself show that the low-light recognition problem has been solved.
Tesla's Robotaxi service page says service hours vary by region and that rides cannot be requested outside operating hours. All that can be confirmed is that the window of availability in Austin has grown by one hour.
What LiDAR can add, and the limits that remain
LiDAR emits its own laser light and uses the light reflected back from objects to measure distance and three-dimensional shape in the surroundings.
That differs from a camera, which reads differences in color and outline between a road and an animal from images captured in visible light. One characteristic of LiDAR is that it can measure distance and shape using its own emitted light even when ambient lighting is poor.
Electrek, which published the original report, viewed Musk's remarks as reflecting a challenge for camera-centered approaches in low-light, low-contrast conditions, and pointed to the difference from LiDAR.
Actual self-driving systems, though, are not necessarily a simple choice between cameras and LiDAR.
Waymo uses a combination of multiple sensor types. In its announcement of the sixth-generation Waymo Driver, the company says the system has 13 cameras, 4 LiDAR units and 6 radar units, with overlapping fields of view used to perceive the surroundings.
In a technical explainer published in August 2026, Waymo separates the roles more clearly.
Cameras read semantic information such as the content of signs and the color of traffic lights, while LiDAR measures the three-dimensional shape of the surroundings. Radar tracks the speed of objects and compensates in conditions that make it hard for cameras to see, such as heavy rain, fog and dust.
The design combines different kinds of information — reading color and meaning, measuring 3D shape, and tracking motion — so that it does not depend on a single sensor.
Still, this is Waymo's own description of its own system. It is not a test result comparing kitten detection or collision-avoidance rates against Tesla on the same roads, in the same lighting, at the same speeds and with the same objects. The type and number of sensors alone cannot be used to rank the overall safety of autonomous driving systems.
LiDAR also has environmental limitations.
Waymo's document on handling fog discusses problems such as water droplets from fog clinging to the surface of optical sensors and shortening the distance the sensors can see. Besides mechanisms to clean the sensors, the company says it is developing ways to recognize the surrounding weather and change how the vehicle drives accordingly.
Using LiDAR can reduce dependence on ambient lighting, but that does not mean the same performance is guaranteed in every weather condition and environment.
What matters is not only which sensor finds an obstacle, but how the system itself recognizes when its perceptible range has narrowed, and what driving decisions it then makes.
Consumer cars have a human who can step in; Robotaxi changes that assumption
In Tesla's published materials, the role expected of users differs greatly between consumer FSD (Supervised) and Robotaxi.
The Model 3 owner's manual requires drivers to pay attention at all times while using FSD (Supervised) and to stay ready to intervene immediately if necessary.
The Robotaxi guide, by contrast, describes a ride experience that does not require a driver. Passengers can call for support or ask the vehicle to stop using the in-car screen or the app.
The engineering analysis of FSD, EA26002, which the U.S. National Highway Traffic Safety Administration (NHTSA) opened on March 18, 2026, is also useful for thinking about this difference.
| Subject | User role / evaluation item described in public documents | What can be read from it |
|---|---|---|
| Consumer FSD (Supervised) | The driver pays attention at all times and intervenes immediately if needed | Not presented as a function that makes the vehicle fully autonomous |
| Robotaxi | A ride experience that does not require a driver; passengers can request a stop or support from the screen or app | Describes how passengers use the service; not a document showing automatic fallback performance when visibility worsens |
| NHTSA's FSD investigation | Evaluates whether the system detects performance degradation from reduced visibility and warns with enough time for the driver to respond | Covers FSD Beta and FSD (Supervised); does not examine Robotaxi's small-animal detection |
This table organizes the public documents available as of October 4, 2026, by user role and investigation scope. It does not show that the software in FSD (Supervised) and Robotaxi is identical, or that their safety performance is the same.
In EA26002, NHTSA covers an estimated 3.2 million vehicles equipped with FSD.
The investigation focuses on whether FSD can detect that its own performance has been degraded by reduced road visibility and warn with enough time left for the driver to respond.
NHTSA describes the FSD in question as an advanced driver-assistance system that relies on cameras and FSD software to detect and react to the road ahead, vehicles, pedestrians, traffic signals and more.
The investigation materials deal with cases in which the system may have failed to properly detect performance degradation when camera visibility worsened, for example because of glare or airborne particles.
However, this is not an investigation that determined a cause or defect regarding the night-time kitten Musk cited. Nor is it a small-animal detection test of Robotaxi itself.
What the investigation shows is that the question of whether obstacles can be recognized needs to be separated from the question of whether the system can tell that its recognition ability has degraded.
With consumer FSD, it is possible to assume that a human driver will step in after the system signals degraded performance.
With a driverless Robotaxi, that assumption does not hold. If the system judges that it cannot perceive its surroundings well enough, the vehicle itself must decide whether to keep going, slow down or stop in a safe place.
A passenger's "Pull Over" is not the same as a driver's immediate intervention
Tesla's Robotaxi also provides a function that lets passengers stop the vehicle themselves.
According to the Robotaxi guide, passengers can select "Pull Over" from the in-car touchscreen or the Robotaxi app. Upon receiving the request, the vehicle stops in a nearby safe location and the passenger can get in touch with Tesla support staff.
But this is a function a passenger uses when they judge it necessary.
It does not directly replace the role of a driver who immediately operates the brakes or steering to avoid an animal that suddenly runs into the road.
The description also does not reveal the conditions under which the vehicle itself would judge that "visibility has worsened and it cannot perceive enough" and automatically begin slowing down or pulling away.
When considering the safety of driverless operation, a stop function that a person can request must be viewed separately from the safety behavior a vehicle performs autonomously after judging danger on its own.
The next focus: can it "find" targets at night, and can it "respond safely"?
In its August 2026 technical document, Waymo explains that, separate from the mechanism that perceives surroundings through sensors, it has an independent onboard system that verifies the driving path chosen by its driving AI.
This verification layer checks the path proposed by the Waymo Driver against physical constraints and traffic rules, and acts to prevent paths that could lead to a collision.
In other words, the design confirms through a separate mechanism not only that the surroundings are recognized correctly, but also that the action chosen from that information is safe.
This too is Waymo's own description of its system. It is not a comparison test with Tesla showing that the challenge of detecting small animals at night has been solved.
Evaluating whether Tesla's Robotaxi can expand night-time operations further requires more than looking at operating hours.
One needs to know how well it can detect hard-to-see obstacles such as small animals, once conditions such as light level, driving speed, object size and distance to the object are made clear. And one needs to know whether, based on that detection, it can slow down or stop at a sufficient distance to avoid a collision.
Equally important is what happens when perception itself degrades.
How does the vehicle judge that it is "not seeing well enough," and in that state, does it keep driving, slow down or stop in a safe place? Musk's post does not reveal specific test results of this kind, or when the challenge might be resolved.
With Austin's end time moved from 10 p.m. to 11 p.m., riders can now hail a Robotaxi for one more hour. But longer operating hours do not mean the same thing as improved night-time recognition.
If night-time operations are to expand further, the question is not simply whether the system can see in the dark. It is whether it can detect hard-to-see objects early enough, and whether it can recognize when it cannot see well enough and switch to safe behavior.