Deep multi-task networks are of particular interest for autonomous driving\nsystems. They can potentially strike an excellent trade-off between predictive\nperformance, hardware constraints and efficient use of information from\nmultiple types of annotations and modalities. However, training such models is\nnon-trivial and requires balancing learning over all tasks as their respective\nlosses display different scales, ranges and dynamics across training. Multiple\ntask weighting methods that adjust the losses in an adaptive way have been\nproposed recently on different datasets and combinations of tasks, making it\ndifficult to compare them. In this work, we review and systematically evaluate\nnine task weighting strategies on common grounds on three automotive datasets\n(KITTI, Cityscapes and WoodScape). We then propose a novel method combining\nevolutionary meta-learning and task-based selective backpropagation, for\ncomputing task weights leading to reliable network training. Our method\noutperforms state-of-the-art methods by a significant margin on a two-task\napplication.\n