二手 Vector 机器人 + GPT-6 Astra 做强化学习进化实验
作者不再手动调参，改成撞到就减权、顺利到达就加权的规则，让装了数字果蝇脑的机器人自己学会绕开椅子和桌腿。
UPDATE: Flyctor is starting to evolve I stopped manually tuning the way it moves. Instead, I gave it a simple rule: Every time it reaches a target without hitting anything, the behavior that got it there is reinforced. Every time it gets stuck, crashes into furniture, or drives in circles, that behavior loses weight. Then I let it run. Flyctor still has the same body: - a used Vector robot from eBay - a camera - GPT-6 Astra turning video into simple sensory signals - a reconstructed fly-brain simulation choosing the next movement But now the connection between "what it sees" and "what it does" is no longer fixed. It changes after every attempt. At first, Flyctor behaved exactly like you would expect from a digital fly brain inside a cheap pet robot. It saw an open path. Turned too early. Corrected too hard. Hit a chair. Repeated the same mistake. Then something changed. It began slowing down before narrow gaps. It stopped making sharp turns near table legs. It started taking longer routes around obstacles instead of repeatedly trying to force its way through them. Nobody wrote those behaviors. Nobody told it what a chair is. The system just learned that some actions lead to movement, and some actions lead to being stuck. I know this is not biological evolution. But watching a tiny robot change its behavior because the old version of itself kept failing feels uncomfortably close. The first Flyctor was a simulation inside a robot body. This one is becoming a slightly better version of itself every time it gets lost.