Running silent background perception and trajectory planning models on consumer vehicles to compare human driver inputs against autonomous decisions.
Real-time divergence metrics calculate the Wasserstein distance between the candidate planner's probabilistic trajectory distribution and the human driver's steering maneuvers.