---
license: gemma
library_name: lerobot
pipeline_tag: robotics
datasets:
- RobotControlStack/duobench
base_model:
- lerobot/pi05_base
tags:
- duobench
- bimanual
- manipulation
- simulation
- pi05
---

# DuoBench π0.5 (sim)

π0.5 policy for the bimanual FR3 Duo setup of [DuoBench](https://arxiv.org/abs/2606.11901), trained on simulation data only for all 11 DuoBench sim tasks (one multi-task policy).
Fine-tuned from [`lerobot/pi05_base`](https://huggingface.co/lerobot/pi05_base) (revision `9e55186`) with [LeRobot](https://github.com/huggingface/lerobot) on the [DuoBench dataset](https://huggingface.co/datasets/RobotControlStack/duobench).

Project page: https://duobench.github.io/
DuoBench Code: https://github.com/RobotControlStack/duobench

## Training
- Data: 50 demonstrations per task from the `<task>/sim` folders of [RobotControlStack/duobench](https://huggingface.co/datasets/RobotControlStack/duobench) (all tasks merged)
- Inputs: `head`, `left_wrist` and `right_wrist` cameras (224x224) and the 16-D state (2 x 7 joints + 2 grippers); output: 16-D joint and gripper action
- 40,000 steps, batch size 16, learning rate 2.5e-5, action chunk 50, seed 1000
- LeRobot commit [`c8ce413d`](https://github.com/huggingface/lerobot/tree/c8ce413d738da15a2eed2d0832315779ea28cbf9); the full configuration is in `train_config.json`

## Results
Success rate (%) in DuoBench simulation over 300 rollouts per task (seeds 0-299) with a Wilson 95% interval.

| Task | Success | 95% CI |
|---|---|---|
| ball_maze | 5.0 | [3.1, 8.1] |
| bin_sort | 1.7 | [0.7, 3.8] |
| block_balance | 0.0 | [0.0, 1.3] |
| carry_pot | 26.0 | [21.4, 31.2] |
| hinge_chest | 26.7 | [22.0, 31.9] |
| join_blocks | 0.0 | [0.0, 1.3] |
| pour_marbles | 0.0 | [0.0, 1.3] |
| spring_door | 19.7 | [15.6, 24.5] |
| transfer_cube | 13.0 | [9.7, 17.3] |
| transfer_gate | 23.3 | [18.9, 28.4] |
| transfer_reorient | 1.3 | [0.5, 3.4] |
| **mean** | **10.6** | |

## Usage
```python
# LeRobot at commit c8ce413d
from lerobot.policies.pi05.modeling_pi05 import PI05Policy

policy = PI05Policy.from_pretrained("RobotControlStack/duobench-pi05-sim")
```
See the [DuoBench code](https://github.com/RobotControlStack/duobench) for the simulation environments and evaluation.

## Limitations
- Trained on simulation data only; not meant for the real robot.
- Loading needs a LeRobot version compatible with the commit above.

## License
This model is a Model Derivative of Gemma (π0.5 is built on [PaliGemma](https://huggingface.co/google/paligemma-3b-pt-224)) and is distributed under the [Gemma Terms of Use](https://ai.google.dev/gemma/terms).
Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms (see `NOTICE`).
Use of this model is subject to the [Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy), which applies to every recipient and downstream user.
The files in this repository are modified: they are the weights of `lerobot/pi05_base` fine-tuned on DuoBench simulation data.

## Citation
If you find this model or DuoBench useful for your work, please consider citing it:
```
@misc{duobench,
  title={{DuoBench}: A Reproducible Benchmark for Bimanual Manipulation in Simulation and the Real World},
  author={Tobias J{\"u}lg and Seongjin Bien and Simon Hilber and Yannik Blei and Pierre Krack and Maximilian Li and Sven Parusel and Rudolf Lioutikov and Florian Walter and Wolfram Burgard},
  year={2026},
  url={https://arxiv.org/abs/2606.11901}
}
```