ManiSkill2: Manipulation Benchmark Demonstrations
ManiSkill2: GPU-accelerated manipulation benchmark with 20+ task environments, expert demonstrations in HDF5 format. Apache 2.0 licensed. RGB, depth, point cloud modalities.
GPU-accelerated manipulation benchmark with 20+ tasks and expert demonstrations. Apache 2.0 licensed.
Apache 2.0 -- Open HDF5 Simulation
Key Stats
| Metric | Value |
|---|---|
| Tasks | 20+ environments (peg insertion, cube stacking, cabinet, drawer, etc.) |
| Robots | Franka Panda, xArm (simulated) |
| Size | ~5 GB demonstrations |
| Format | HDF5 |
| Modalities | RGB, depth, point cloud, joint state, actions |
| License | Apache 2.0 |
| Related datasets | 99 ManiSkill-related datasets on HuggingFace |
What is ManiSkill2?
ManiSkill2 is a GPU-accelerated manipulation simulation benchmark from UCSD's HAO Lab (Hao Su group). It provides a diverse set of manipulation environments with standardized evaluation protocols, plus expert demonstrations that serve as training data for imitation learning research.
Key advantages include GPU parallelization (thousands of environments in parallel on a single GPU), rich observation spaces (RGB + depth + point cloud), and a growing ecosystem of 99 related datasets on HuggingFace. ManiSkill3, the latest iteration, further improves simulation speed and adds soft-body manipulation tasks.