UniSim provides backend-neutral physics contracts and optional engine
adapters for robot learning and simulation. The PyPI distribution is
unisim-core; the Python import namespace is unisim.
A single SimBackend contract covers state access, control, reset, and
domain-randomization boundaries, so the same task code runs on MuJoCo,
Motrix, Drake, MJWarp, Genesis, IsaacGym, or IsaacSim without engine-specific
branches. The base install depends only on NumPy; every engine SDK is an
optional extra loaded lazily, and importing unisim never imports an engine.
UniSim is the extracted, backend-neutral physics layer used by UniLab. UniLab retains Hydra configuration, task/env/manager lifecycle, robot assets, RL training, checkpoints, and sim2sim policy I/O; UniSim owns the physics contract, adapter lifecycle and state translation, optional-runtime diagnostics, and the shared subprocess IPC layer. There is exactly one production implementation of each backend, owned by this repository — UniLab only assembles task-owned scene and configuration inputs and consumes the public contract. UniSim never imports UniLab.
pip install unisim-core # base: contract, factory, fake backend
pip install "unisim-core[mujoco]" # plus an engine extra when neededAvailable extras: mujoco, motrix, drake, mjwarp, genesis, newton,
isaacgym, isaacsim, superdex. SuperDex's Python 3.12 CPU development
profile is described in docs/superdex.md. The Isaac extras
are empty spellings because those
vendor SDKs are not redistributable; their adapters discover dedicated worker
installations at construction time. See
docs/support-matrix.md for the full adapter
support matrix.
The mujoco, mjwarp, and newton extras share the MuJoCo 3.11 /
MuJoCo-Warp 3.11 / warp-lang 1.16.0 line and can be installed together in one
environment. newton keeps exact pins (newton==1.5.1 with its coupled
runtimes), while mjwarp follows the 3.11 line with mujoco-warp~=3.11.0.
The public boundary is deliberately lazy and safe to import anywhere:
from unisim import SimBackend, create_backendConstruct a backend through the factory with a package-neutral SceneCfg:
backend = create_backend("mujoco", scene=scene_cfg, num_envs=64, sim_dt=0.01)
backend.materialize() # cold path: parse XML, build engine objects
backend.reset()
state = backend.get_state()
backend.step(ctrl) # hot path: validated arrays, cached handlesEach adapter fails closed with an actionable, backend-specific diagnostic when its optional runtime is missing — no backend is silently downgraded to another engine. Engine-native model and data objects never escape the adapter; state and control flow through validated NumPy arrays.
A deterministic FakeBackend and the assert_backend_conformance helper let
consumers test task code without any engine installed. BenchmarkCase and
BenchmarkResult are reserved schema extension points for a future benchmark
package; no workload runner is implemented here.
External worker roots can be configured with UNISIM_ISAACGYM_HOME,
UNISIM_ISAACGYM_PYTHON, UNISIM_ISAACSIM_HOME, and
UNISIM_ISAACSIM_PYTHON. The package also accepts the former UNILAB_*
spellings as a migration fallback.
docs/architecture.md— ownership boundary between UniSim and UniLabdocs/support-matrix.md— adapter install and runtime requirementsdocs/migration.md— migrating from the historical UniLab backend layerdocs/mocap-reset-contract.md— selected-world mocap poses and MJWarp primitive/contact/joint randomizationdocs/benchmark-api.md— reserved benchmark schemasdocs/release.md— TestPyPI and automated production release procedure
make sync # locked environment plus the MuJoCo test extra
make check # Ruff + pytest
make package # source distribution and wheel for local inspectionEvery adapter must document its supported Python/platform/runtime matrix and pass the conformance helper before it is published.
If UniSim contributes to your research, please cite the UniLab paper:
@article{jia2026unilab,
title = {UniLab: A Heterogeneous Architecture for Robot RL Beyond
GPU-Dominant Paradigms},
author = {Yufei Jia and Zhanxiang Cao and Mingrui Yu and Heng Zhang and
Shenyu Chen and Dixuan Jiang and Meng Li and Xiaofan Li and
Yiyang Liu and Junzhe Wu and Zheng Li and XiLin Fang and
Tingyu Cui and Shengcheng Fu and Haoyang Li and Anqi Wang and
Zifan Wang and Dongjie Zhu and Chenyu Cao and Zhenbiao Huang and
Ziang Zheng and Jie Lu and Xin Ma and Zhengyang Wei and
Xiang Zhao and Tianyue Zhan and Ye He and Yuxiang Chen and
Yizhou Jiang and Yue Li and Haizhou Ge and Yuhang Dong and
Fan Jia and Ziheng Zhang and Meng Zhang and Xiwa Deng and
Zhixing Chen and Hanyang Shao and Chenxin Dong and Yixuan Li and
Yizhi Chen and Bokui Chen and Kaifeng Zhang and Hanqing Cui and
Yusen Qin and Ruqi Huang and Lei Han and Tiancai Wang and
Xiang Li and Yue Gao and Guyue Zhou},
journal = {arXiv preprint arXiv:2605.30313},
year = {2026},
url = {https://arxiv.org/abs/2605.30313}
}When you use a specific backend through UniSim, please also cite the
corresponding engine. The mujoco and drake adapters build on the MuJoCoUni
and DrakeUni runtimes, so cite those alongside the original engines:
% MuJoCo
@inproceedings{todorov2012mujoco,
title = {MuJoCo: A Physics Engine for Model-Based Control},
author = {Todorov, Emanuel and Erez, Tom and Tassa, Yuval},
booktitle = {2012 IEEE/RSJ International Conference on Intelligent Robots and Systems},
pages = {5026--5033},
year = {2012},
doi = {10.1109/IROS.2012.6386109}
}
% MuJoCoUni (runtime of the `mujoco` adapter)
@article{jia2026mujocouni,
title = {MuJoCoUni: Persistent Batched Runtime Primitives for MuJoCo},
author = {Jia, Yufei and Wu, Junzhe},
journal = {arXiv preprint arXiv:2605.24922},
year = {2026}
}
% MotrixSim
@software{motrixsim2026,
title = {MotrixSim: A Physics Simulation Engine for Robotics and Embodied AI},
author = {{Motphys Team}},
year = {2026},
url = {https://motrixsim.readthedocs.io/},
note = {Python binary package}
}
% Drake
@misc{tedrake2019drake,
title = {Drake: Model-Based Design and Verification for Robotics},
author = {Russ Tedrake and the Drake Development Team},
year = {2019},
url = {https://drake.mit.edu}
}
% DrakeUni (runtime of the `drake` adapter)
@software{drakeuni,
title = {DrakeUni: Experimental Drake Batch Simulation Runtime for UniLab},
author = {{UniLab Team}},
year = {2026},
url = {https://pypi.org/project/drake-uni/},
note = {Python binary package}
}
% MJWarp
@software{mujoco_warp,
title = {MuJoCo Warp: A GPU-Accelerated MuJoCo Backend},
author = {{Google DeepMind}},
year = {2025},
url = {https://github.com/google-deepmind/mujoco_warp}
}
% Newton
@software{newton2025,
title = {Newton: GPU-accelerated physics simulation for robotics and
simulation research},
author = {{Newton Contributors}},
year = {2025},
url = {https://github.com/newton-physics/newton}
}
% Genesis
@misc{genesis,
title = {Genesis: A Universal and Generative Physics Engine for Robotics
and Beyond},
author = {Genesis Authors},
month = {December},
year = {2024},
url = {https://github.com/Genesis-Embodied-AI/Genesis}
}
% Isaac Gym
@inproceedings{makoviychuk2021isaacgym,
title = {Isaac Gym: High Performance GPU-Based Physics Simulation for
Robot Learning},
author = {Makoviychuk, Viktor and Wawrzyniak, Lukasz and Guo, Yunrong and
Lu, Michelle and Storey, Kier and Macklin, Miles and
Hoeller, David and Rudin, Nikita and Allshire, Arthur and
Handa, Ankur and State, Gavriel},
booktitle = {Proceedings of the Neural Information Processing Systems Track
on Datasets and Benchmarks},
year = {2021}
}
% Isaac Sim
@software{nvidia2022isaacsim,
title = {NVIDIA Isaac Sim},
author = {{NVIDIA}},
year = {2022},
url = {https://developer.nvidia.com/isaac/sim}
}Apache-2.0, see LICENSE.