---
library_name: robo-orchard-lab
base_model: Qwen/Qwen3.5-2B
tags:
  - robotics
  - vision-language-action
  - simulation
  - roboorchardsim
---

# HoloBrain v0.5 sim: RoboOrchardSim six tasks finetune

This repository provides a HoloBrain 0.5 action-model checkpoint and its configuration for six RoboOrchardSim tasks. The implementation is in [RoboOrchardLab `projects/holobrain_0_5`](https://github.com/HorizonRobotics/RoboOrchardLab/tree/master/projects/holobrain_0_5). The checkpoint is under [`robo_orchard_sim_contest_six_tasks/`](./tree/main/robo_orchard_sim_contest_six_tasks); it uses a separate [Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B) backbone.

## Intended use

This checkpoint is intended for research and evaluation on the six simulation tasks represented by its bundled configs: `place_a2b_000`, `spatial_pick_000`, `pick_color_000`, `stack_cubes_000`, `joint_direction_000`, and `open_drawer_000`. Use the RoboOrchardLab HoloBrain 0.5 implementation to load the action model, processor, and task-specific settings. This package has not been documented as a general-purpose chat model or validated for physical-robot deployment.

## Download

Download this repository, then run the backbone command **from inside** the model directory:

```bash
hf download HorizonRobotics/HoloBrain_v0.5_sim --local-dir .
cd robo_orchard_sim_contest_six_tasks
hf download Qwen/Qwen3.5-2B \
  --revision 15852e8c16360a2fea060d615a32b45270f8a8fc \
  --local-dir ckpt/Qwen3.5-2B
```

`model.config.json` expects the backbone at `ckpt/Qwen3.5-2B`. That directory was a local symlink in the source package and is not distributed here. The pinned Qwen revision matches the backbone used with this export. See the [directory README](./blob/main/robo_orchard_sim_contest_six_tasks/README.md) for the short download instruction.

## Package contents

```text
robo_orchard_sim_contest_six_tasks/
├── README.md
├── model.safetensors
├── model.config.json
├── panda_droid_rigid_inference.config.json
├── panda_droid_rigid_processor.json
├── training_config.json
├── training_result.json
├── configs/
│   ├── __init__.py
│   ├── config_holobrain_0_5.py
│   ├── data_configs/
│   │   ├── __init__.py
│   │   ├── config_issac_dataset.py
│   │   ├── config_libero_dataset.py
│   │   └── config_robotwin_dataset.py
│   ├── dataset_factory.py
│   ├── dataset_specs.py
│   ├── deploy_specs.py
│   ├── panda_droid_smoke.json
│   ├── panda_droid_train.json
│   ├── robot_specs.py
│   └── transforms.py
└── urdf/
    └── franka_droid.pruned.urdf
```

## Evaluation and limitations

No aggregate success rates or benchmark scores are included with this model package, so this card makes no performance claim. Its intended task distribution is defined by the six simulation task configs above. The Qwen backbone must be downloaded separately, and successful use also depends on the matching RoboOrchardLab code and simulation environment.