本文介绍了 RoboBrain 2.0，这是我们最新一代的具身视觉-语言基础模型，旨在统一物理环境中复杂具身任务的感知、推理和规划。它有两种变体：轻量级 7B 模型和全尺寸 32B 模型，采用异构架构，包括视觉编码器和语言模型。尽管尺寸紧凑，RoboBrain 2.0 在广泛的具身推理任务中表现出色。在空间和时间基准测试中，32B 变体取得了领先的结果，超越了之前的开源和专有模型。特别是，它支持关键的具身 AI 能力，包括空间理解（如可操作性预测、空间引用、轨迹预测）和时间决策（如闭环交互、多智能体长时规划、场景图更新）。本报告详细介绍了模型架构、数据构建、多阶段训练策略、基础设施和实际应用。我们希望 RoboBrain 2.0 能够推动具身 AI 研究，并成为构建通用具身智能体的实际一步。代码、检查点和基准测试可在 https://superrobobrain.github.io 获取。
We introduce RoboBrain 2.0, our latest generation of embodied vision-language foundation models, 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, surpassing prior open-source and proprietary models. In particular, it supports key real-world embodied AI 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 scene graph updating). 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. The code, checkpoint and benchmark are available at https://superrobobrain.github.io.
核心贡献 · Key contributions
提出 RoboBrain 2.0，一种统一感知、推理与规划的具身视觉语言基础模型。 Introduces RoboBrain 2.0, a new embodied vision-language foundation model unifying perception, reasoning, and planning.
异构架构，包含视觉编码器和语言模型，提供 7B 和 32B 两种版本。 Heterogeneous architecture with vision encoder and language model, in 7B and 32B variants.
构建大规模多样化数据集，涵盖空间推理、时间建模和因果推理。 Large-scale diverse datasets synthesized for spatial reasoning, temporal modeling, and causal inference.
在多个具身空间和时间基准上取得领先结果，超越此前模型。 Achieves state-of-the-art results on multiple embodied spatial and temporal benchmarks, outperforming prior models.
开源代码、检查点和基准，推动具身人工智能研究。 Open-source release of code, checkpoints, and benchmarks to advance embodied AI research.
局限 · Limitations
依赖合成数据可能限制对真实新场景的泛化能力。 Reliance on synthetic data may limit generalization to novel real-world scenarios.
在 VSI-Bench 和 EmbSpatial 等基准上未达到最佳性能，仍落后于一些专有模型。 Not top-performing on all benchmarks (e.g., VSI-Bench, EmbSpatial), trailing some proprietary models.
思维链推理为可选功能，可能在复杂任务中产生不准确或冗余的解释。 Chain-of-thought reasoning is optional and may produce inaccurate or verbose explanations in complex tasks.
缺乏对真实机器人硬件的广泛评估，结果主要来自模拟基准。 Lack of extensive evaluation on real robotic hardware; results mainly from simulated benchmarks.
训练数据主要来自室内环境（如 CA-1M、AI2Thor），可能对室外或极端场景泛化不足。 Training data biased toward indoor scenes (e.g., CA-1M, AI2Thor), limiting generalization to outdoor or extreme environments.
论文章节 · Sections（共 44）
摘要Abstract
RoboBrain 2.0 技术报告RoboBrain 2.0 Technical Report
1 引言1 Introduction
2 架构2 Architecture
2.1 输入模态与词元化2.1 Input Modalities and Tokenization
2.2 视觉编码器与投影2.2 Vision Encoder and Projection
2.3 大语言模型解码器与输出表示2.3 LLM Decoder and Output Representations