What happens to an AI's prediction when the camera shifts or rotates?
Instead of testing images and hoping for coverage, we can now mathematically bound these predictions—a rigorous evaluation step required to certify safety-critical applications for autonomous vehicles.
The Joby Superpilot™ team is presenting at
#CVPR2026 in Denver this week alongside our
@safe_intel_ai colleagues. By modeling camera motion through homographies and applying Lipschitz optimization, we derive provably tight bounds on how pixels change, no simulation or 3D scene model required.
We're proud to report up to 89% faster speeds, 7% tighter bounds, and present the formal verification of a vision-based landing system under 3D camera motion perturbations.
Watch the full presentation ⬇️