An overhaul could help the system see better, but there are challenges.
Confirming earlier statements by Elon Musk, Tesla Motors announced today that it will move towards using radar as the primary sensor for its’ Autopilot systems, using advanced signal processing to build a robust computer model of cars’ surroundings. Up to this point, radar only supplemented Autopilot’s camera-based detection system.
The changes will begin with Version 8 of Tesla’s vehicle software, which will include more robust radar modeling and begin gathering data to improve radar performance—but will not immediately change how Autopilot sees and makes decisions.
Get Data Sheet, Fortune’s technology newsletter.
The shift to radar comes in the wake of the first-ever fatality of an operator using the Autopilot system. The circumstances of the May crash, as well as some similar but nonfatal incidents, suggested that the camera-based Autopilot system may have problems detecting overhanging objects. Final findings of an NTSB investigation of that crash have not been released.
While making radar the primary sensor for Autopilot decision-making could help prevent similar incidents, the company says it presents its own challenges. While radar is good at seeing through environmental obstructions like smoke and rain, it is not as good as a camera at detecting people, and has serious problems seeing wood or plastic objects.
Most challenging of all, metallic objects appear highly reflective to a radar system. This, the company says, is a particular problem when it comes to concave surfaces. The bottom of a soda can on the road, for example, “can appear to be a large and dangerous object” in radar. To avoid those kinds of false positives, the updated Tesla system will assemble much more information about objects than before, and construct a moving 3D “point cloud” that evaluates objects over time, not just as a single snapshot.
For more on autonomous vehicles, watch our video.
Cars will not immediately begin seeing or behaving differently. Instead, all Tesla cars will begin gathering information using the new software, and uploading it to a central database. This will help the company construct a “whitelist” of objects like road signs and bridges that could appear as threats on radar, but are actually safe.
The system will not begin acting on radar-based decisions until, according to Tesla, “the data shows that false braking events would be rare.” Radar will then be used to supplement camera-based decisions, and slowly take over more decision-making as the system demonstrates its reliability.
“The net effect of this [change],” Tesla says, “Combined with the fact that radar sees through most visual obscuration, is that the car should almost always hit the brakes correctly even if a UFO were to land on the freeway in zero visibility conditions.”
Today’s announcement also included an array of other Autopilot tweaks and upgrades, including better reaction to braking from leading cars, and better avoidance of nearby cars nearing the edge of their lane.