Stone Tao
Stone Tao
They look nice and we can probably simulate+render it faster than their code with the GPU sim (habitat rearrange task has so many objects and still simulates fairly fast)
Running PPO for this task on commit d01141da8a1f9460dbee6c2d223e1a02dbf46698 gets the original speed (converged policy gets around 4k FPS on 4090) Running it on commit dba1ca4ffa2271542ccc66f3e755f728240c305b gets the slower speed (converged...
Things like CEM should be easy to make parallelized which can help generate demonstrations / verify rewards https://github.com/google-deepmind/mujoco_mpc/blob/main/docs/OVERVIEW.md is a nice reference
Currently you can only use one GPU environment to do replay. There is also some problems with object geometry randomizations which are cpu/gpu seed dependent that need to be addressed...
Currently they use some pretty expensive/large complex meshes for just a table that is then ultimately greenscreened out, it can be simplified a lot. Once simplified the speed should massively...
ManiSkill2 with 16 parallel envs can simulate somewhere in the 2-3k region FPS, but in ManiSkill3 with the CPU backend it drops down to around 1-2K FPS.
Support an option to render at different frequencies. If render mode is human, then we internally call the render_human function at the appropriate frequency. Otherwise we accumulate render image data...