Researchers Train a Rat Named Freddy to Play a Simplified GTA Via Brain-Computer Interface
Sad_Measurement_7502 · reddit · 2026-10-20
- 玩法：作者用基于真实啮齿类 BCI 实验的闭环模拟系统，让老鼠 Freddy 通过多通道神经接口「玩」极简版 GTA。自适应解码器把脑活动翻译成 5 个动作：前进、左转、右转、交互、加速，不含瞄准和完整物理。
- 奖励设计：上车 1 份奖励、完成短途驾驶 2 份、长途 3.5 份、成功躲避检测 2.5 份，闲置有小惩罚；奖励即时发放。同时监测心率与应激水平，压力过高会降低有效奖励，防止乱按。
- 结果：Freddy 无视开放世界，专刷回报最高的循环；只有短任务收益下降时才做长任务。解码器会在线学习哪些神经模式能换来奖励，越玩越熟练。
- 开源：自适应解码器与奖励塑形代码已上传 GitHub（Rat-Theft-Auto）。
More from Fun
- Synthesia and NVIDIA debut a conversational avatar of cartoon Jensen Huang at GTC Berlin — alexvoica · 2026-10-21
- Amanita Design slams AI slop animation using Machinarium assets — Promptmethus · 2026-10-21
- Kevin Roose and Casey Newton's NPR podcast 'Machine Gods' premieres on a new set — kevinroose · 2026-10-21
- MiMo 2.6's reasoning trace is just 'Hmm' everywhere, and one Redditor can't stop hearing it — gaspoweredcat · 2026-10-21
- Developer gave his Grok bot $1,000 to build itself a body — parts keep arriving — Baconbrix · 2026-10-21
- Cloudflare's Matthew Prince takes a subtle jab at AWS during Connect keynote — fforres · 2026-10-21