Real2Sim policies

Three pi0.5 policy checkpoints evaluated by Real2Sim, which runs openpi pi0.5 policies on a Franka FR3 arm with the bimanual YUBI gripper in the MuJoCo simulator, against photorealistic 3D Gaussian Splatting backgrounds. Weights only: no optimiser or training state, and no tokenizer (openpi downloads the PaliGemma tokenizer on first use).

Paper: arXiv:2610.17521
Code: the Real2Sim repository (link to follow)
Scenes and meshes: airoa-org/yubi-real2sim-assets

Folder	Policy	Training step	Profile token	Release profiles
`pi05_box_bs/`	box packing, baseline	80,000	`box80k`	12: every rig, room and task
`pi05_box_ct/`	box packing, contact-dimension variant	80,000	`cd80k`	8: every rig and room, box tasks
`pi05_cup/`	cup on plate	4,000	`cup1trim`	4: every rig and room, cup task

Each folder contains:

File	Content
`params/`	Model weights (Orbax OCDBT checkpoint)
`assets/<name>/norm_stats.json`	Action/state normalisation statistics read by openpi at inference time
`prompt.yaml`	Task instructions the policy was trained with (`pi05_box_bs`, `pi05_box_ct`)
`_CHECKPOINT_METADATA`	Orbax checkpoint metadata
`LICENSE`	Copy of the Gemma Terms of Use
`GEMMA_PROHIBITED_USE_POLICY.txt`	Copy of the Gemma Prohibited Use Policy
`LICENSE-APACHE-2.0`	Apache License 2.0 (licence of openpi and of its pi0.5 model)
`NOTICE`	Required Gemma notice, use restrictions, modification notice and attribution
`README.txt`	Description of the folder

Download

# one policy (about 12 GB)
hf download airoa-org/yubi-real2sim-policies --include "pi05_cup/*" --local-dir yubi-real2sim-policies

# all three (about 37 GB)
hf download airoa-org/yubi-real2sim-policies --local-dir yubi-real2sim-policies

The Real2Sim installer (scripts/download_public_assets.py --profile <profile>) downloads only the policies and scenes a profile needs and places them under $REAL2SIM_INPUTS_ROOT/checkpoints/<name>.

Origin

pi0.5 vision-language-action policies fine-tuned by AIRoA with openpi from Physical Intelligence's pi0.5 model, on AIRoA robot demonstration data recorded with the YUBI gripper.

License

The weights contain a PaliGemma model, so each checkpoint is a Model Derivative of Gemma and is distributed under the Gemma Terms of Use (LICENSE in each folder). Use is subject to the Gemma Prohibited Use Policy. The pi0.5 model and openpi are licensed by Physical Intelligence under the Apache License 2.0 (LICENSE-APACHE-2.0). See NOTICE in each folder.

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms.

Citation

@article{nishizawa2026yubi,
  title={YUBI-Real2Sim: 3DGS Scene Verification for Reliable Policy Evaluation},
  author={Nishizawa, Hiroki and Ohkawa, Takehiko and Scianca, Nicola and Matsushima, Tatsuya and Morishima, Shigeo and Ota, Kei},
  journal={arXiv preprint arXiv:2610.17521},
  year={2026}
}