v3.0.0b5 Released

[StoneT2000]

ManiSkill v3.0.0b5 is released! Upgrade by running

pip install --upgrade mani_skill

Major Changes:

• Several new tasks added, including a nice GPU parallelized PushT environment developed by a community member (Xander Hinrichsen). All are solved by PPO in minutes.
• A simple parallel GUI render mechanism has been added (similar to how isaac renders environments) although note it is only useful for cool demo videos
• Quadruped task has improved reward function for nicer locomotion behaviors
• GPU parallelized Soft Actor Critic (SAC) Baseline added
• Reverse forward curriculum learning (RFCL) baseline added, a fast online imitation learning algorithm (not GPU parallelized but is still the fastest learning from demos algorithm provided at the moment)
• More demonstration datasets have been added sourced from motion planning solutions, reinforcement learning, and simple teleoperation for a few tasks.
• New CPU/GPU contact forces API that is much easier to work with and internally highly optimized so you don't need to write the complex GPU facing code yourself, see: https://maniskill.readthedocs.io/en/latest/user_guide/tutorials/custom_tasks/advanced.html#contact-forces-on-the-gpu-cpu
• Added AI2THOR Scenes (without lighting so ray-tracing renders will look weird but default rendering (shader="default") will look correct).
• A few tasks and robots have been optimized so state simulation is faster now
• tool for converting pd_joint_pos or pd_joint_delta_pos control to pd_ee_delta_pose control has been fixed and is now usable

Release notes here: https://github.com/haosulab/ManiSkill/releases/tag/v3.0.0b5
Some demos of the new work:

anymalc_parallel.mp4

PushT-v1_rt.mp4

opencabinet_parallel.mp4

pickcube_parallel.mp4