Untitled

주목할 점:

• Memory based control (recurrent LSTM)을 로봇에 적용한 첫 사례
• Dynamic randomization을 적용해서 RNN이 실제 로봇에 적용했을 때 발생하는 overfitting 문제를 해결.

Existing work:

• simple memoryless network architectures

In this paper:

• recurrent neural networks (RNNs) for sim-to-real biped locomotion, allowing for policies that learn to use internal memory

• RNNs are found to outperform memoryless policies in simulation but not on the real biped due to overfitting to the simulation physics

  ⇒ use dynamics randomization in training to prevent overfitting

Contents:

• Method
• Algorithm
• Results