July
@july
If you think of reality as a simulator (now you've got me thinking @aviationdoctor.eth) - you essentially have to represent the entire world as some sort of convex objects, that are running at discrete time (since we can't do continuous time in any sort of the types of computers / machines we have today. They'd also need to be programmed to have rigid bodies, collision detection etc - and handle friction etc. Also photo realistic sensor simulation i.e. sensor simulation of your perception is pretty expensive. We don't even really have the compute to run a full world physics engine, we have to use stuff like ray-tracing to only show what you see, and then on top of it, we do rasterization to increase the speed in which we can show what we see. I think that's why "domain randomization" is interesting, because if the model (our RL policy) sees enough simulated variation - then transferring sim2real -> moving the policy to the real world works well (treat the world like a simulator) - that's interesting
1 reply
1 recast
23 reactions
agusti
@bleu.eth
nvidia dropped some huge models for real world sim environments to train ai for robots in
1 reply
0 recast
3 reactions
