Roadmap#
We are constantly working to improve ManiSkill on all aspects, from simulation/rendering speed to usability. Below is a brief list of planned upcoming features that will come to ManiSkill in the future
GPU parallelized low-level mobile manipulation tasks (at the moment you can just take random actions but there are no reward/success metrics)
More fully interactive scene datasets (currently there is only ReplicaCAD. There is also AI2THOR scenes but they are not interactable)
More industrial robots properly modelled and added into simulation (Google Robot, UR5, Unitree H1 etc.)
Larger demonstration datasets, especially tele-operated demonstrations
RL Baselines (Algorithms like SAC, TDMPC2 etc.)
Offline RL Baselines (Algorithms like IQL, CQL, etc.)
Behavior Cloning Baselines (Algorithms like 3D Diffusion Policy, etc.)
Vision Language Model Baselines (Algorithms like PerAct, VoxPoser, etc.)
Easy to use domain randomization tools (randomizing lighting, camera poses, object textures etc.)
Wider variety of verified tasks
A new, fast and accurate soft body simulator
Simple beautification tools to make high quality renders/recordings of trajectories in simulation
Object database of high quality, simlation ready assets (rigid-body, soft-body, articulations) for use in your tasks
More sensor types like Lidar
Real2Sim evaluation pipelines