Logo RoboBrain 2.0: See Better. Think Harder. Do Smarter.

BAAI RoboBrain Team
Technical Report Code Checkpoints

We are excited to introduce RoboBrain2.0, the most powerful open-source embodied brain model to date. Compared to its predecessor, RoboBrain 1.0, our latest version are designed to unify perception, reasoning, and planning for complex embodied tasks in physical environments. It comes in two variants: a lightweight 7B model and a full-scale 32B model, featuring a heterogeneous architecture with a vision encoder and a language model. Despite its compact size, RoboBrain 2.0 achieves strong performance across a wide spectrum of embodied reasoning tasks. On both spatial and temporal benchmarks, the 32B variant achieves leading results in most cases, surpassing prior open-source and proprietary models. In particular, it supports key real-world embodied intelligence capabilities, including spatial understanding (e.g., affordance prediction, spatial referring, trajectory forecasting) and temporal decision-making (e.g., closed-loop interaction, multi-agent long-horizon planning, and real-time scene memory). This report details the model architecture, data construction, multi-stage training strategies, infrastructure and practical applications. We hope RoboBrain 2.0 advances embodied AI research and serves as a practical step toward building generalist embodied agents.

Performance

Benchmark results
Task capabilities
Benchmark comparison across spatial reasoning and temporal task planning. RoboBrain2.0-32B achieves state-of-the-art (SOTA) or near-SOTA performance on nine spatial reasoning benchmarks and three temporal reasoning benchmarks. It not only outperforms leading open-source models such as Cosmos-Reason1 and Qwen2.5-VL, but also surpasses closed-source models like Gemini 2.5 Pro, o4-mini and Claude Sonnet 4.
Task capabilities
Benchmark comparison across spatial reasoning and temporal task planning. RoboBrain2.0-32B achieves state-of-the-art (SOTA) or near-SOTA performance on nine spatial reasoning benchmarks and three temporal reasoning benchmarks. It not only outperforms leading open-source models such as Cosmos-Reason1 and Qwen2.5-VL, but also surpasses closed-source models like Gemini 2.5 Pro, o4-mini and Claude Sonnet 4.
Task capabilities
Benchmark comparison across spatial reasoning and temporal task planning. RoboBrain2.0-32B achieves state-of-the-art (SOTA) or near-SOTA performance on nine spatial reasoning benchmarks and three temporal reasoning benchmarks. It not only outperforms leading open-source models such as Cosmos-Reason1 and Qwen2.5-VL, but also surpasses closed-source models like Gemini 2.5 Pro, o4-mini and Claude Sonnet 4.

Architecture

teaser
Architecture of RoboBrain2.0. The model supports multi-image, long video, and high-resolution visual inputs, along with complex task instructions and structured scene graphs on the language side. Visual inputs are processed via a Vision Encoder and MLP Projector, while textual inputs are tokenized into a unified token stream. All inputs are fed into a LLM Decoder that performs long-chain-of-thought reasoning and outputs structured plans, spatial relations, and both relative and absolute coordinates.

Task Capabilities

teaser
Task Capabilities of RoboBrain2.0. RoboBrain2.0 supports interactive reasoning with long-horizon planning and closed-loop feedback, spatial perception for precise point and bbox prediction from complex instructions, temporal perception for future trajectory estimation, and scene reasoning through real-time structured memory construction and update.

highlight

Demo Videos

System Stability

This video demonstrates the model's referential ability in color recognition and its stability in continuous operation.

Real-time Scene Adaptation

This video demonstrate the model's rapid scene adaptation ability and its capability to judge object proximity, recognize orientation, and determine distance.

Real-time Voice Interruption Adjustment

This video demonstrates the model's capabilities in object spatial relationship recognition, multi-step reasoning, rapid interactive reasoning, and real-time interruption adjustment.

Part-level Orientation-related Referring

This video demonstrates the model's capabilities in object spatial height recognition and part-level orientation-related region identification.

Functionality-oriented Referring

This video demonstrating the model's capabilities in object spatial height recognition and illuminated area identification.

Multi-step Spatial Referring with Reasoning

This video demonstrates the model's object spatial relationship recognition and multi-step spaital referring with reasoning capability.

Structured Arrangement

This video demonstrates the model's ability to understand spatial relationships and pattern reasoning between objects.

Mobile Manipulation

This video demonstrates the model's ability to control a humanoid for both tabletop object manipulation and indoor navigation.

Object Attribute Recognition

This video demonstrates the model's ability to accurately recognize and differentiate objects by their sizes and its stability in continuous operation.

Object Affordance Localization

This video demonstrates the model's capability in object affordance prediction (grasping the handle of the mug) as well as locating objects based on their colors and distances.

Spatial Relations Reasoning

This video demonstrates the model's spatial reasoning capabilities, including distance perception (nearest), position awareness (left and front), and free space localization.

Spatial Referencing and Vacancy Detection

This video demonstrates the model's object referencing capability based on spatial relations and its ability to locate vacant areas in 3D space.

Training & Evaluation

We highlight the distributed training framework FlagScale developed by BAAI Framework R&D team, and the evaluation framework FlagEvalMM developed by BAAI FlagEval team. Both are used for RoboBrain 2.0. Many thanks to the teams for their contributions!

flagscale
FlagScale is a distributed training framework designed for large-scale models, supporting efficient training and evaluation of models like RoboBrain 2.0. It provides a flexible and scalable solution for training large models across multiple GPUs and nodes.
flageval
FlagEvalMM is a comprehensive evaluation framework for multi-modal models, including RoboBrain 2.0. It provides a suite of benchmarks and metrics to assess the performance of multi-modal models in various tasks, ensuring robust evaluation and comparison.

Citation

If you find our model helpful, feel free to cite it: 

@article{RoboBrain 2.0 Technical Report,
    title={RoboBrain 2.0 Technical Report},
    author={BAAI RoboBrain Team},
    journal={arXiv preprint arXiv:TODO},
    year={2025}
    }
@article{RoboBrain1.0,
    title={Robobrain: A unified brain model for robotic manipulation from abstract to concrete},
    author={Ji, Yuheng and Tan, Huajie and Shi, Jiayu and Hao, Xiaoshuai and Zhang, Yuan and Zhang, Hengyuan and Wang, Pengwei and Zhao, Mengdi and Mu, Yao and An, Pengju and others},
    journal={arXiv preprint arXiv:2502.21257},
    year={2025}
}
@article{RoboOS,
    title={RoboOS: A Hierarchical Embodied Framework for Cross-Embodiment and Multi-Agent Collaboration},
    author={Tan, Huajie and Hao, Xiaoshuai and Lin, Minglan and Wang, Pengwei and Lyu, Yaoxu and Cao, Mingyu and Wang, Zhongyuan and Zhang, Shanghang},
    journal={arXiv preprint arXiv:2505.03673},
    year={2025}
}
@article{zhou2025roborefer,
    title={RoboRefer: Towards Spatial Referring with Reasoning in Vision-Language Models for Robotics},
    author={Zhou, Enshen and An, Jingkun and Chi, Cheng and Han, Yi and Rong, Shanyu and Zhang, Chi and Wang, Pengwei and Wang, Zhongyuan and Huang, Tiejun and Sheng, Lu and others},
    journal={arXiv preprint arXiv:2506.04308},
    year={2025}
}
@article{Reason-RFT,
    title={Reason-rft: Reinforcement fine-tuning for visual reasoning},
    author={Tan, Huajie and Ji, Yuheng and Hao, Xiaoshuai and Lin, Minglan and Wang, Pengwei and Wang, Zhongyuan and Zhang, Shanghang},
    journal={arXiv preprint arXiv:2503.20752},
    year={2025}
}
@article{Code-as-Monitor,
    title={Code-as-Monitor: Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection},
    author={Zhou, Enshen and Su, Qi and Chi, Cheng and Zhang, Zhizheng and Wang, Zhongyuan and Huang, Tiejun and Sheng, Lu and Wang, He},
    journal={arXiv preprint arXiv:2412.04455},
    year={2024}
}