Domain Adaptation

Sim2Real

The “sim2real” problem refers to the challenge of transferring knowledge and skills learned in simulation environments to the real world.

  • this problem exists arises because simulations can never fully capture the complexity and variability of the real world

Idea: Learn to race in a video game, and translate those skills in real life.

Ways to bridge the Sim2Real gap (source):

  1. System identification: build a mathematical model for a physical system; (this is the simulator in the context of RL)
  2. Domain Adaptation (adversarially generated environments)
  3. Domain Randomization

Both DA and DR are unsupervised. Compared to DA which requires a decent amount of real data samples to capture the distribution, DR may need only a little or no real data. So Domain Randomization might be better.

repo NVIDIA “Towards Optimal Strategies for Training Self-Driving Perception Models in Simulation” at NeurIPS 2021:

Synthetic Data Generation at WATonomous

Ideas:

This I don’t know how I got this. I was looking through BEVFusion, and then found LSS which the paper mentioned, which was written by Jonah Philion. And then I stalked his advisor who is prof. Sanja Fidler.

But the professor Sanja Fidler, I was browsing through her research, and found this Meta-Sim: Learning to Generate Synthetic Datasets