Part 3/9:
Advanced Benchmarking and Performance Metrics
Apple reported that their trained driving policy outperformed previous state-of-the-art benchmarks while never being exposed to human driver data during training. The results demonstrated a significant improvement in robustness, with agents averaging 17.5 years of driving between incidents during trials in simulated environments.
Minimalistic Training Approach
Using a sparse reward function, the training process required only basic rules—such as avoiding collisions and following traffic laws—to develop effective driving policies. The minimalist approach differentiates Apple's methods from those that require extensive human data input, suggesting a shift towards more automated forms of training in AI.