具身智能中的sim2real的gap是什么？

● Sim2Real gap - The domain gap between simulated data and real world data

● Models trained in simulation without proper configurations fail in the real world

● The appearance gap is the inability to make simulated images exactly replicate what the real world looks like

● Until simulators become completely photorealistic there will always be an appearance gap

● Luckily with progression in simulators this gap is quickly closing

● Simulation mimics a limited set of scenes, not necessarily reflecting the diversity and distribution of objects of those captured in the real world

● The real world is composed of many different objects and structures that can be hard to recreate in simulation

• Domain Adaptation - Techniques to transfer the source domain (simulation) to the target domain (real world)

• Domain Randomization - Create diverse randomized simulation variations where reality is seen as just another variation by the DNN

• System Identification - Create an accurate simulation that matches the properties of reality

参考文献：

Structured Domain Randomization - https://arxiv.org/abs/1810.10093
Metasim - https://arxiv.org/abs/1904.11621
Metasim 2 - https://arxiv.org/abs/2008.09092
Deception net - https://arxiv.org/pdf/1904.02750.pdf
Slide 12 Graphic - https://lilianweng.github.io/lil-log/2019/05/05/domain-randomization.html