Finding the sweet spot in unknown environments 100× faster

When a system needs to find its best operating point in an unknown environment—like a vehicle optimizing fuel consumption—it faces a classic dilemma: try new settings to learn (exploration) or stick with what works (exploitation). The authors discovered that dual control problems have a hidden mathematical structure: a convex outer layer wrapping a nonlinear inner map. By exploiting this geometry, they built a solver that computes the next action in ~83 microseconds on real vehicle hardware, roughly 10× faster than existing methods. This matters for robotics and autonomous systems where decisions must happen in milliseconds.