ACE-Brain/ACE-Brain-0.5-8B

| 🤗 HuggingFace  | 📔 Technical Report  | 🤖 Github  |

Overview

ACE-Brain-0.5 is a unified embodied foundation model for Physical Agentic AI. It extends ACE-Brain-0 from an understanding-centric spatial model into a closed-loop embodied model that can perceive the physical world, plan under goals, act through robot bodies, monitor execution progress, and improve from accumulated experience.

ACE-Brain-0.5 organizes robot intelligence into five tightly coupled cognitive functions: Spatial Perception, Decision Making, Embodied Interaction, Self Monitoring, and Self Improvement. A single 8B backbone instantiates the core perception-planning-action-evaluation loop, supporting object and affordance grounding, 3D and egocentric spatial reasoning, long-horizon task planning, navigation and manipulation action generation, and progress estimation for verification and recovery.

Key Features

• Unified Embodied Foundation Model: Organizes robot intelligence into a single closed-loop model spanning Spatial Perception, Decision Making, Embodied Interaction, Self Monitoring, and Self Improvement.
• SSR+ training paradigm: Extends Scaffold-Specialize-Reconcile with a Reactivate stage, combining task-vector merging with targeted fine-tuning to unify spatial reasoning, grounding, navigation, manipulation, and progress estimation without cross-task interference.

Inference Example

from transformers import Qwen3VLForConditionalGeneration, AutoProcessor

model = Qwen3VLForConditionalGeneration.from_pretrained(
    "ACE-Brain/ACE-Brain-0.5-8B",
    dtype="auto",
    device_map="auto",
)

processor = AutoProcessor.from_pretrained("ACE-Brain/ACE-Brain-0.5-8B")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
            },
            {"type": "text", "text": "Describe this image."},
        ],
    }
]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
)
inputs = inputs.to(model.device)

generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False,
)
print(output_text)

Citation

If you find ACE-Brain-0.5 useful for your research and applications, please consider citing our technical report.

@misc{brainteam2026acebrain05unifiedembodiedfoundational,
      title={ACE-Brain-0.5: A Unified Embodied Foundational Model for Physical Agentic AI},
      author={Brain Team and Ziyang Gong and Haoming Gu and Zehang Luo and Tianyi Zhang and Tao Tao and Yixiao Chi and Zhe Liu and Lingsi Zhu and Jingyuan Liu and Anke Tang and Songze Li and Yilun Kong and Ningjing Liu and Tianyu Zhu and Yunpeng Qing and Shuang Luo and Xiang Liu and Shi Fu and Dawei Nie and Sixiang Liu and Zhexi Wen and Feng Pan and Xiaofeng Wang and Zhi Hou and Chunxiao Liu and Xue Yang and Junchi Yan and Hengshuang Zhao and Dacheng Tao and Xiaogang Wang},
      year={2026},
      eprint={2607.04426},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2607.04426},
}