7 comments

[ 2.8 ms ] story [ 29.4 ms ] thread
Images don't look that realistic, they have a marked video-gamy look. I wonder how well stuff learned in this environment transfers to the real world.
I was going to argue the opposite that such high fidelity images are not really necessary to learn basic behaviour, in fact they look like overkill to me
Yeah image quality is irrelevant for training... but physics do. I was hoping they would have ultra-realistic physics engine.

Like have X amount of force applied to a ball in simulation and real-world and then watch as the simulation perfectly matches how the real world ball bounces around. That would be impressive. Instead it doesn't seem as they're adding that much more to a game engine other than a couple robot characters.

As it says in the first sentence, this simulator is built on top of Unreal Engine 4.
There are a few papers which shows that this works quite well.
I read the article but I don’t understand what the end goal is here. Are they training game AI? Can someone help me understand what this is for?
Synthetic training data. Sometimes it's cheaper / safer / easier to work with simulated data rather than real world sensor data. For some types of machine learning, a simulation is "real enough".

Someone asked me to build a startup doing the same. His most convincing use case was training camera systems to detect people falling overboard on cruise ships, etc. It's impractical / dangerous to use real world camera data for that.

There are other advantages, e.g., being able to generate simulations where everything is already labeled.