AI Learns To Drive

Before judging the driver in AI Learns To Drive, decide what you rewarded it for. The official description lets you mix reward functions for different goals, including speed, caution and drifting. Its evolutionary training then uses the best agents as parents for the next generation.

That makes the objective part of the construction. The fastest-looking run and the behavior you asked the system to favor are not necessarily the same question. The listing does not publish the reward formulas, so it cannot establish which weights reliably produce a particular driving style. It does establish that the player can change the incentives rather than only watch a fixed training demonstration.

Information and control are configurable too. Vision rays, speed, acceleration and wheel angle are named inputs; steering, throttle, brakes and battery boost are named outputs. Campaign challenges unlock additional inputs and outputs, so availability is part of the progression rather than an assumption that every setting is present at the start.

One negative English report posted October 4 says the game was good but reports no progress after 20,000 generations. It recorded 700 minutes at review; the network, reward settings and build are not specified. That cannot establish a universal failure or explain the plateau, but it qualifies the idea that simply running more generations guarantees improvement.

The useful comparison is between what a network can sense, what it can command and what earns it a better score. Changing one of those gives you a reason to compare the resulting driving behavior. A collision is an outcome to explain within that setup, not evidence that evolutionary training has promised a universally competent driver.