Genesis – a generative physics engine for general-purpose robotics

Given what's there today, especially the sizzle reel, I'm pretty dubious.

If the author drops an amazing generative text-to-sim system on top of this... THAT is impressive - but effectively orthogonal to what's there - so I'm withholding excitement for now.

Take the time to read over the repo. It is not revolutionary. It is an integration of a bunch of third party packages (which are largely C/C++ libraries with Python wrappers, not "pure python"!). The stuff unique to Genesis is adequate implementations of well-known techniques, or integration code.

The backflip is awesome but plausibly explained by the third party RL library, and they include an example program which... runs a third party library to do just this.

The performance numbers are so far beyond real world numbers as to be incoherent. If you redefine what all the words mean, then the claims are not comparable to existing claims using the same words. 43 million FPS means, if my math is right, you are spending 70 clocks per frame on a 3ghz processor. On a 4080 you would have ~500k clocks in the same period, but that implies 100% utilization with zero overhead from Amdahl's law. (Also, Hi, Erwin, maybe you think these claims are 100% realistic for meaningful workloads in which case I'll gladly eat crow since I have a huge amount of respect for Bullet!)

I can only judge what's released now, not a theoretical future release, and what's here now is something a really good developer could bang out in a couple of months. The USP is the really good spin around the idea that it's uniquely suited for AI to produce beyond-SOTA results.

I have slightly longer form thoughts: https://x.com/bengarney/status/1869803238389887016

This is kind of correct. For the demos they show those will simulate quite slowly due to the soft body / fluid simulation. Their current code does not parallelize soft body/fluid simulation like they do with rigid bodies. Running their fluid sim code now with default settings (which are tuned for speed, not accuracy) gives maybe max 10-100x real time speed (their rigid body sim is magnitudes faster but definitely not faster than existing GPU simulators like Isaac or MJX). Rendering at that quality will also take forever but everyone in this area of research always runs the slow high quality ray tracing pipelines for the demo videos (which run in the 10-100 FPS range depending on settings). Another note is they claim realism but this is only visual realism, there doesn’t exist soft body sim solvers that transfer to the real world accurately enough yet for eg robotics. They are good enough for visual data generation which will be pretty cool.

(full disclosure I work on ManiSkill/SAPIEN which runs on PhysX from NVIDIA, Genesis authors claim they are essentially faster than PhysX by 10-80x).