Autonomous driving systems are only as good as the scenarios they train on. A new preprint on arXiv argues that the distribution of those scenarios is a key factor in how well a vehicle learns to drive. The authors point out that existing approaches typically construct or adapt training distributions using surrogate criteria—heuristics that stand in for actual driving difficulty or value.
To move beyond those heuristics, the paper outlines an approach inspired by Lamarckian evolution, in which training strategies themselves are discovered through evolutionary competition. Rather than manually designing scenario distributions, the method appears to let competing strategies evolve over time, though the abstract provided does not detail the mechanism or results.
Because the abstract is truncated, the full method and experimental outcomes are not available from the source. Still, the framing highlights a growing interest in treating the training curriculum itself as an optimisation problem.