Excitation-Aware Friction Capability Bounds for Uncertainty-Aware Speed Planning of Autonomous Ground Vehicles

Speed planners must respect the tire–road friction limit, but friction is identifiable only when the tires work near saturation. We show that, for a saturating tire model, the Fisher information about friction in a measurement scales with the sixth power of tire utilisation, so ordinary driving leaves friction essentially unobserved and an estimate can remain unchanged long after the road has changed. Rather than trusting an estimate because it once converged, we propose an information-aware trust model: the friction used for planning is an approximate lower confidence bound whose width follows the information recently accumulated, relaxing toward a conservative prior when excitation is absent and with distance from the last informative excitation. A forward–backward friction-circle planner uses this value with a conditional feasibility guarantee. In simulation (20-run Monte Carlo, ±10% stiffness mismatch) a hidden friction drop caused lane departures in every run of certainty-equivalent planning and none with the proposed method, which recovered 71% of the time lost by fixed conservative planning; a fixed safety margin on the estimate was either unsafe or slower, depending on how slippery the road became. In CARLA the method had no departures in 10 runs, whereas fixed-nominal planning departed in every run, certainty-equivalent planning in 17 of 18 and a fixed safety margin in 3 of 10, while being 2.6% faster than fixed conservative planning. We report the limits openly: the bound is approximate, is overconfident when the tire stiffness is underestimated, and in CARLA safety relied chiefly on the distance-based trust decay.Preprint; under review at IEEE Robotics and Automation Letters. Code, data and video: https://github.com/munavirzaman-git/friction-bound