physical-superintelligence-lab/SIMPLE

SIMPLE – Simulation‑Based Policy Learning and Evaluation

What it is – SIMPLE is an open‑source simulation platform for training and benchmarking whole‑body humanoid locomotion‑manipulation policies. It bundles a high‑fidelity physics stack (Isaac Sim 4.5 + MuJoCo 3.3) with a large library of assets (1000+ Objaverse objects, 50+ Habitat scenes) and a suite of 50+ humanoid tasks (e.g., carry‑box, pick‑and‑place, door opening). The repo ships the environment code, data‑generation pipelines, and evaluation scripts that work with the foundation model Ψ₀ (Psi‑0).

Key components

• Multi‑robot support – Franka, Aloha bimanual arms, Dex‑Mate wheeled robot, and Unitree G1 humanoid.
• Rich content – Thousands of 3D objects and dozens of indoor scenes for diverse training data.
• Task catalog – Whole‑body loco‑manipulation benchmarks, each with three OOD difficulty levels (visual distractors, lighting changes, spatial pose perturbations).
• Client‑server evaluation – A decoupled architecture where the Ψ₀ model runs as an inference server (Psi‑0 repo) and SIMPLE runs the simulation client, enabling fast policy rollout and video logging.
• Data pipeline – Scripts for tele‑operation collection, automated motion‑planning generation, post‑processing, and fine‑tuning of policies.
• Extensible – Supports integration of World‑Action Models such as Cosmos3 and DreamZero.

Installation options

1. UV (quickest) – Install system deps, uv package manager, then run uv sync --all-groups. Includes a helper script to install the CuRobo planner.
2. Nix – Recommended for fresh Linux hosts; detailed guide in the docs.
3. Docker – Pre‑built images for both the SIMPLE client and the Ψ₀ server are provided on GitHub Container Registry.

Getting started (example)

# Pull Docker images
docker pull ghcr.io/physical-superintelligence-lab/psi0:latest   # model server
docker pull ghcr.io/physical-superintelligence-lab/simple:latest # simulation client

# Run the server (Psi‑0) – replace $RUN with your checkpoint tag
docker compose run --rm serve-psi0-sonic-http \
    --policy psi0 --port 8014 --ckpt-step 40000 \
    --run-dir .runs/finetune/$RUN

# In another terminal, run the SIMPLE client for a whole‑body carry‑box task
GPUs=1 docker compose run --rm eval-sonic-wbc \
    simple/G1WholebodyXMoveBendCarryBoxSonic-v0 psi0 \
    --data-dir data/simple/G1WholebodyXMoveBendCarryBoxSonic-v0/dr-level-0 \
    --host 127.0.0.1 --port 8014 \
    --episode-start 0 --num-episodes 5 --dr-level 0 \
    --eval-dir data/eval/sonic-psi0

The command launches five evaluation episodes and writes per‑episode videos under data/eval/….

Benchmark results (excerpt) – On six core tasks, Ψ₀ achieves near‑perfect success at Level 0 (10/10) and degrades gracefully to Level 2, outperforming baselines such as GR00T, OpenPi, and DreamZero. Full tables are in the README.

Documentation & resources

• Full docs: https://psi-lab.ai/SIMPLE/docs/
• Model checkpoint: Hugging Face USC-PSI-Lab/psi-model
• Dataset: Hugging  Face USC-PSI-Lab/psi-data
• Paper (arXiv 2606.08278): https://arxiv.org/abs/2606.08278

License – MIT (see LICENSE).

Citation – If you use SIMPLE or the Ψ₀ model, cite the two arXiv papers listed in the README.