Contributors: Songlin Wei, Zhenhao Ni, Jie Liu, Zhenyu Zhao and more (to appear) ...

What is SIMPLE?

SIMPLE stands for SIMulation-based Policy Learning and Evaluation.

It is a simple simulation environment supports:

• multiple agents: (franka arm/aloha bimanual arms/dexmate wheeled robot and unitree g1 humanoid!)
• 1000+ Objaverse assets
• 50+ Habitat HSSD scenes
• 50+ humanoid wholebody loco-manipulation tasks

System Requirements

SIMPLE is built on top of IsaacSim 4.5 and MuJoCo 3.3.

Component	Minimum	Recommended
OS	Ubuntu 22.04	Ubuntu 22.04
CPU	Intel Core i7 / AMD Ryzen 7	Intel Core i9 / AMD Ryzen 9
RAM	32 GB	64 GB
GPU	NVIDIA RTX 2070 (8 GB VRAM)	NVIDIA RTX 3080 Ti / 4090 (16+ GB VRAM)
NVIDIA Driver	535.x	Latest
CUDA	12.x	12.x
Python	3.10	3.10
Storage	50 GB SSD	100+ GB NVMe SSD

An RTX-class NVIDIA GPU is required. GTX and older architectures are not supported.

Installation

Clone the project:

git clone git@github.com:physical-superintelligence-lab/SIMPLE.git

Change directory to the project root:

cd SIMPLE

Pull all submodules

git submodule update --init --recursive

We offer three options for setting up SIMPLE:

[Option 1] UV setup (Quickest)

Install uv if not already done

curl -LsSf https://astral.sh/uv/install.sh | sh

Install all dependencies at once

UV_HTTP_TIMEOUT=3000 bash scripts/setup_python_env.sh

Install CuRobo

bash scripts/install_curobo.sh

Activate the environment:

source .venv/bin/activate

Verify the installation by printing the version number

python -c "import simple; print(simple.__version__)"

[Optional] Build the docs.

make live

Open http://127.0.0.1:8005 in a browser to view the documentation.

The document are working in progress. Feel free to raise questions using github issue, we will try to complete the document construction as soon as possible.

[Option 2] robo-nix setup

SIMPLE keeps Python dependencies in uv, but uses robo-nix for the native robotics runtime: CUDA, graphics, MuJoCo, Isaac Sim runtime libraries, media libraries, and native build tools.

Install robo:

curl -fsSL https://raw.githubusercontent.com/ausbxuse/robo-nix/develop/scripts/install.sh | sh

Prepare the runtime, enter it, and install Python dependencies:

robo up
robo shell
bash scripts/setup_python_env.sh
bash scripts/install_curobo.sh

Verify the runtime:

python -c "import simple; print(simple.__version__)"
bash scripts/tests/check_datagen.sh
bash scripts/tests/check_eval.sh

On hosts where the Nix daemon does not trust the nixpkgs-python binary cache, robo up fails before compiling CPython from source. Prefer adding the cache to the daemon trust list. If you intentionally accept the slow source-build fallback, run:

ROBO_NIX_ALLOW_SOURCE_PYTHON=1 robo up

Isaac Sim 4.5 is not stable with every NVIDIA driver branch. NVIDIA driver 595 has been observed to segfault during Isaac SimulationApp startup on an RTX 5090 Laptop GPU, while drivers 575 and 580 have both been tested and worked. If datagen crashes right after Isaac/RTX startup, try a 575- or 580-series driver and rerun robo up after removing .robo-nix/host-graphics, .robo-nix/shell-env, and .robo-nix/shell-env.key.

See robo-nix setup for notes on the project runtime contract.

[Option 3] Docker setup

We also support building and running SIMPLE in docker. Please refer to the documents for docker setup.

📊 Simulation Benchmarking Results

This is a preliminary benchmark with 6 tasks accompanying the Psi-0 project. Please also checkout Psi-0 for more details of intergrating Psi-0 with SIMPLE.

To rigorously evaluate the robustness and generalization of the learned policies, we design three evaluation levels with progressive out-of-distribution variations applied to the training environment:

The evaluation environments are provided in the huggingface repository USC-PSI-Lab/psi-data.

• Level 0 (Visual & Distractors): Randomizes table materials and the types/initial positions of distractor objects.
• Level 1 (Lighting): Includes Level 0 variations + extreme changes in lighting conditions.
• Level 2 (Spatial pose): Includes Level 1 variations + perturbations to the initial positions of the target objects.

Success rates are reported out of 10 evaluation trials per level (Level 0 | Level 1 | Level 2).

Baseline / Task	G1Wholebody XMove PickTeleop-v0	G1Wholebody BendPickMP-v0	G1Wholebody Handover Teleop-v0	G1Wholebody Locomotion PickBetweenTables Teleop-v0	G1Wholebody Tabletop GraspMP-v0	G1Wholebody XMove BendPick Teleop-v0
Psi0	10 | 10 | 6	10 | 10 | 10	7 | 7 | 10	7 | 5 | 6	10 | 10 | 8	10 | 9 | 9
GR00T N1.6	10 | 10 | 7	7 | 7 | 6	1 | 3 | 3	0 | 0 | 0	9 | 9 | 7	4 | 4 | 1
OpenPi π0.5	7 | 5 | 1	10 | 10 | 8	5 | 4 | 5	3 | 3 | 3	10 | 10 | 8	0 | 0 | 0
InternVLA-M1	0 | 0 | 0	5 | 5 | 0	0 | 0 | 0	0 | 0 | 0	0 | 0 | 0	3 | 5 | 7
H-RDT	0 | 0 | 2	0 | 0 | 1	0 | 1 | 0	0 | 0 | 0	0 | 0 | 0	0 | 0 | 0
DreamZero	- | - | -	- | - | -	- | - | -	- | - | -	9 | 10 | 10	- | - | -
EgoVLA	0 | 1 | 2	7 | 5 | 8	0 | 4 | 3	0 | 0 | 0	10 | 10 | 7	3 | 5 | 4
Diff. Policy	3 | 3 | 2	10 | 8 | 6	3 | 2 | 4	4 | 0 | 0	8 | 9 | 8	0 | 0 | 0
ACT	10 | 9 | 6	10 | 9 | 9	4 | 4 | 6	6 | 5 | 7	10 | 10 | 8	6 | 8 | 8

Citation

@misc{wei2026psi0,
  title={$\Psi_0$: An Open Foundation Model Towards Universal Humanoid Loco-Manipulation}, 
  author={Songlin Wei and Hongyi Jing and Boqian Li and Zhenyu Zhao and Jiageng Mao and Zhenhao Ni and Sicheng He and Jie Liu and Xiawei Liu and Kaidi Kang and Sheng Zang and Weiduo Yuan and Marco Pavone and Di Huang and Yue Wang},
  year={2026},
  eprint={2603.12263},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2603.12263}, 
}

License

This project is licensed under the MIT.

See the LICENSE file for details.