模仿学习Imitation Learning
Imitation Learning
让机器照着示范学动作，而不是自己从头试错。Let a machine learn actions from demonstrations instead of trial-and-error from scratch.
模仿学习指从专家示范中学策略：记录「什么状态下做什么动作」的成对样本，再训练模型复现这些动作。
最直接的实现是行为克隆，把示范当监督学习来拟合；示范没覆盖到的状态容易出错，DAgger 等方法让专家在学习者走过的状态上继续标注来补救。
它与强化学习互补：一个靠现成示范，一个靠奖励试错；机器人遥操作采集的演示数据常走这条路。
Imitation learning learns a policy from expert demonstrations: record paired samples of what action to take in what state, then train a model to reproduce those actions.
The most direct implementation is behaviour cloning, fitting demonstrations as supervised learning; states not covered by demonstrations are error-prone, and methods such as DAgger ask experts to label states visited by the learner.
It complements reinforcement learning: one uses ready demonstrations, the other uses reward trial-and-error; demonstration data collected by robot teleoperation often takes this path.
机器人学动作最常见的起点就是它：遥操作采来的示范直接变成训练数据。分清「从示范学」和「从奖励学」，能看懂为什么很多机器人项目先花钱采数据，而不是先设计奖励函数。This is the most common starting point for robot action learning: teleoperated demonstrations become training data. Separating learning from demonstrations and learning from reward explains why many robot projects spend on data collection before designing reward functions.
回到名词地图看第1站的其他 35 个词 →Back to the map: the other 35 terms in station 1 →