ManiSkill2
Simulated benchmark of 20 manipulation task families, rigid and soft body, with 2,000+ objects and 4M+ demo frames.1
Comparisons with real robots
Tried on real robots, not compared1
Details
One point-cloud RL policy for PickCube: 91.0% success in simulation, 60.0% over 50 real trials (ROKAE xMate3Pro arm, Robotiq 2F-140, RealSense D415). Pinch: same motion-planned action sequence run in sim and real, compared qualitatively. Measured by the ManiSkill2 authors; no correlation statistic.
Details
About
- What it is
- Benchmark Inferred4
More
Classified by the Atlas from how the authors describe and distribute it.
- Built by
- UC San Diego; Tsinghua University1
More
15 authors; Hao Su last author.
- Released
- 2022-085
More
First release tag v0.2.0 on 2022-08-15; arXiv v1 2023-02-09.
- Version
- 0.5.36
- Last update
- v0.5.3 released 2023-09-22 (RNG seeding bug fix)5
More
ManiSkill3 README: original ManiSkill2 code lives at the v0.5.3 tag.
- Status
- Superseded6
Setup
- Runs in
- Simulation1
More
SAPIEN rigid-body simulation plus custom GPU MPM (Warp) for soft bodies.
- Robot
- One arm, Arm on wheels, Two arms1
More
'stationary/mobile-base, single/dual-arm'.
- Robot model
- panda, mobile_panda, xmate3 robot definitions Inferred7
More
Files in mani_skill2/agents/robots at v0.5.3.
- Setting
- Tabletop, Whole home Inferred1
More
Tabletop rigid/soft tasks plus household articulated-object tasks; mapping is ours.
- Size
- 20 task families; 2,000+ object models; 4M+ demonstration frames1
More
Same numbers on project site https://maniskill2.github.io/.
Scoring and access
- Scored by
- Success rate1
More
Appendix: e.g. PickCube success = cube within 2.5 cm of goal and robot static.
- Trials
- 100 episodes per task with varied initial states (most rigid tasks)1
More
Appendix task protocols.
- Who runs it
- The organisers run the tests1
More
'we implement a cloud-based evaluation system'.
- Leaderboard
- Official leaderboard1
More
https://sapien.ucsd.edu/challenges/maniskill/2022/ refused connection; project site invites to the challenge without results.
- Code licence
- Apache-2.0 for rigid-body environments; soft-body environments follow NVIDIA Source Code License for Warp3
More
README 'License' section at tag v0.5.3; repo LICENSE file is Apache-2.0.
- Data licence
- Assets: CC BY-NC 4.0 (README). Demonstrations: Hugging Face card haosulab/ManiSkill2 says apache-2.08
More
Two primary sources give different licences for different artefacts; record both.
- Commercial use
- Unclear Inferred3
More
Code permissive, assets non-commercial, soft-body code under Warp licence. Not legal advice.
Sources 8
- 1ManiSkill2: A Unified Benchmark for Generalizable Manipulation SkillsPaper · Feb 2023 · checked 10 Oct 2026
- 2Semantic Scholar API recordIndex · checked 10 Oct 2026
- 3mani-skill/ManiSkill on GitHub (file README.md)Repository · checked 10 Oct 2026
- 4ManiSkill2: A Unified Benchmark for Generalizable Manipulation SkillsPaper · Feb 2023 · checked 10 Oct 2026
- 5mani-skill/ManiSkill on GitHub (releases)Repository · checked 10 Oct 2026
- 6mani-skill/ManiSkill on GitHub (file README.md)Repository · checked 10 Oct 2026
- 7mani-skill/ManiSkill on GitHub (file robots?ref=v0.5.3)Repository · checked 10 Oct 2026
- 8haosulab/ManiSkill2 on Hugging Face (dataset)Repository · checked 10 Oct 2026
Change history
- Created as a basic entry: identity facts checked at primary sources (phase 1 re-verification).