ManiSkill v3.0.0b16 Released

[StoneT2000]

Overview

The latest version brings a number of optimizations and upgrades!

Highlights

Camera sensors are now lazily initialized which enables some massive GPU speed/memory savings, you can run more environments in parallel with rendering on now, especially for RL training. Depending on situation you might be saving up to 8GB of GPU memory.

Moreover we have overhauled our trajectory replay tool to now more seamlessly support CPU/GPU replay of CPU/GPU trajectories. Using environment state you can also replay a trajectory collected on one machine on another machine with a different number of parallel environments. Data generation with visual data is now also far faster as well.

We have also added more baseline scripts for running various baselines and updated our wandb dashboard with more results. Moreover the Action Chunking Transformer behavior cloning baseline has now been added.

Breaking Changes

• All demonstration data are labelled in their filename or metadata with what sim-backend they were collected on. Previously this used to be either cpu or gpu. As we are considering supporting more physics engines, we have renamed these as physx_cpu and physx_cuda. Huggingface demo datasets have all been updated to reflect this.
• Not a big breaking change but the default hyperparameter for the ppo state and rgb baselines finite_horizon_gae is now default False, which trains faster in general. We still recommend using ppo_fast.py for fast state based training (which by default has finite_horizon_gae False as well).

Full release notes here: https://github.com/haosulab/ManiSkill/releases/tag/v3.0.0b16