Notes on Implicit Behavioral Cloning

Key idea

Behavioral cloning policies are often represented by explicit continuous feed-forward models that mappingg directly from input observations $o\in \mathcal{O}$ to output action $a\in \mathcal{A}$.

$$\hat{a}=F_{\theta }(o)$$

In this work, they reformulate BC using the composition of argmin with an energy function $E$ to represent the policy,

$$\hat{a}=\arg \underset{a}{\min }{E}_{\theta }(o,a)$$

Then train the model with different EBM training methods.

What you can learn?

1. From explicit to implicit. If there’s a way to transfer the methods/optimization goal from explicit to implicit, there’s a way to write a paper.

Drawbacks

Not found.