Galaxea Open-World Dataset and G0 Dual-System VLA Model

01 动机Motivation

VLA 模型的发展面临一个核心瓶颈：缺乏大规模、高质量的开放世界机器人数据。现有数据集（如 BridgeData V2、DROID、Open X-Embodiment）大多在受控或人工布置的实验室场景中采集，场景多样性有限、语言标注粗糙，难以支撑 VLA 向真实世界的泛化。The development of VLA models faces one core bottleneck: the scarcity of large-scale, high-quality open-world robot data. Existing datasets (such as BridgeData V2, DROID and Open X-Embodiment) are mostly collected in controlled or artificially arranged laboratory scenes, with limited scene diversity and coarse language annotation, which makes it hard to support the generalization of VLA to the real world.

"a substantial bottleneck persists due to the scarcity of large-scale, high-quality, open-world robot data."

Galaxea 开放世界数据集与 G0 双系统框架概览 — **Figure 1**：Galaxea Open-World Dataset 在真实人居与工作环境中收集高质量机器人行为数据；G0 是一个由 VLM（慢系统，负责规划）与 VLA（快系统，负责执行）组成的双系统框架。**Figure 1**: the Galaxea Open-World Dataset collects high-quality robot behaviour data in real human living and working environments; G0 is a dual-system framework composed of a VLM (the slow system, in charge of planning) and a VLA (the fast system, in charge of execution).

500h高保真数据总时长Total hours of high-fidelity data

100K演示轨迹数量Number of demonstration trajectories

150+任务类别Task categories

50真实场景数量Number of real-world scenes

现有多具身数据集（如 OXE）虽然规模庞大，但来自不同机器人平台，动作空间不统一，反而可能因"具身鸿沟（embodiment gap）"损害特定机器人的学习效果。本文以单一机器人平台（Galaxea R1 Lite）在 11 个真实地点的 50 个场景采集数据，涵盖居民区、餐饮、零售和办公等多种环境，并提供细粒度子任务级别的语言标注，从根本上解决数据多样性与一致性的矛盾。Existing multi-embodiment datasets (such as OXE) are large in scale, but they come from different robot platforms with inconsistent action spaces, and may in fact hurt learning on a specific robot because of the "embodiment gap". This paper collects data with a single robot platform (Galaxea R1 Lite) across 50 scenes at 11 real-world locations, covering residential, dining, retail and office environments, and provides fine-grained subtask-level language annotation, which fundamentally resolves the tension between data diversity and consistency.

Galaxea R1 Lite 机器人平台与真实场景数据采集 — **Figure 2**：**(a)** Galaxea R1 Lite 平台：23-DoF 双臂移动机器人（6-DoF 双臂 + 3-DoF 躯干 + 6-DoF 全向底盘），最高速度 1.5 m/s，承载 5 kg，配备立体 RGB 头部相机及双腕 Intel RealSense D405 RGB-D 相机。**(b)** 机器人编队在多个真实世界环境中采集数据，确保数据的高保真度与场景多样性。**Figure 2**: **(a)** the Galaxea R1 Lite platform: a 23-DoF dual-arm mobile robot (6-DoF dual arms + 3-DoF torso + 6-DoF omnidirectional base), top speed 1.5 m/s, payload 5 kg, equipped with a stereo RGB head camera and dual-wrist Intel RealSense D405 RGB-D cameras. **(b)** a fleet of robots collects data across multiple real-world environments, ensuring high data fidelity and scene diversity.

数据采集三原则Three Principles of Data Collection

Observability：视觉流包含所有任务相关信息Observability: the visual stream contains all task-relevant information
数量与质量：简单任务需约 100 条演示；复杂序列质量优先Quantity and quality: a simple task needs about 100 demonstrations; for complex sequences, quality comes first
Linguistic grounding：子任务级语言标注，支持多模态对齐Linguistic grounding: subtask-level language annotation, supporting multimodal alignment

数据集核心统计Core Dataset Statistics

1,600+ 种独特物体1,600+ distinct objects
58 种操作技能（从 "pick" 到 "whole-body" 协调）58 manipulation skills (from "pick" to "whole-body" coordination)
11 个物理采集地点11 physical collection sites
使用同构遥操作（isomorphic teleoperation）采集Collected with isomorphic teleoperation

02 方法Method

G0 由两个异步运行的系统组成：G0-VLM（System 2，慢系统）负责高层规划，将自然语言指令分解为子任务序列；G0-VLA（System 1，快系统）以 flow matching 方式生成 action chunk，实现精细执行。整个框架采用三阶段渐进式训练。G0 consists of two asynchronously running systems: G0-VLM (System 2, the slow system) handles high-level planning and decomposes a natural-language instruction into a subtask sequence; G0-VLA (System 1, the fast system) generates action chunks by flow matching for fine-grained execution. The whole framework adopts three-stage progressive training.

**Figure 7**：G0-VLA 架构与三阶段训练流程。**Stage 1** 在跨具身数据（约 1,000 小时 OXE + 500 小时 Galaxea + 200 小时内部数据，共 1,700 小时）上以自回归方式预训练 VLM（使用 FAST tokenizer 将连续动作转为离散索引）。**Stage 2** 在具有子任务级语言标注的 Galaxea 数据上训练完整 VLA（预训练 VLM + 新初始化的 Action Transformer），以 flow matching loss 监督动作重建。**Post-train** 用最多 100 条轨迹对下游任务进行微调。颜色标识不同 token 模态。**Figure 7**: the G0-VLA architecture and the three-stage training pipeline. **Stage 1** pre-trains the VLM autoregressively on cross-embodiment data (about 1,000 hours of OXE + 500 hours of Galaxea + 200 hours of in-house data, 1,700 hours in total), using the FAST tokenizer to turn continuous actions into discrete indices. **Stage 2** trains the complete VLA (the pre-trained VLM + a newly initialized Action Transformer) on Galaxea data with subtask-level language annotation, supervising action reconstruction with a flow matching loss. **Post-train** fine-tunes on a downstream task with at most 100 trajectories. Colours mark the different token modalities.

G0-VLA：快系统（System 1）G0-VLA: the Fast System (System 1)

G0-VLA 以预训练 VLM（基于 PaLiGemma，含 SigLIP 视觉编码器 + 单层 MLP 投影 + Transformer）为骨干，新增 Action Transformer（flow matching action expert）作为动作生成头。给定语言指令、视觉观测与本体感知状态，生成 action chunk A_t = a_t:t+k（水平为 k）。G0-VLA takes a pre-trained VLM (based on PaLiGemma, with a SigLIP vision encoder + a single-layer MLP projection + a Transformer) as its backbone and adds an Action Transformer (flow matching action expert) as the action generation head. Given the language instruction, visual observations and proprioceptive state, it generates an action chunk A_t = a_t:t+k (horizon k).

Stage 1：仅训练 VLM，以 FAST tokenizer + 标准 cross-entropy loss 预测下一个动作 token；使用高层任务描述（Galaxea 数据中仅用高层描述）。Stage 1: train the VLM only, predicting the next action token with a FAST tokenizer + a standard cross-entropy loss; high-level task descriptions are used (only high-level descriptions are used for the Galaxea data).
Stage 2：两大关键要素——单一具身（action space 一致）+ 语言-动作对齐（子任务级细粒度 language-action pairs）；以 flow matching loss 训练完整 VLA。Stage 2: two key ingredients — a single embodiment (consistent action space) + language-action alignment (fine-grained subtask-level language-action pairs); the full VLA is trained with a flow matching loss.
Post-train：在下游任务上微调，最多使用 100 条轨迹，训练目标与 Stage 2 相同。Post-train: fine-tuning on downstream tasks with at most 100 trajectories; the training objective is the same as in Stage 2.

G0-VLM：慢系统（System 2）G0-VLM: the Slow System (System 2)

G0-VLM 基于开源 Qwen2.5-VL 进行指令微调，融合 Galaxea 数据中的人工标注子任务与合成高层指令。训练时对关键帧（子任务终止或夹爪状态变化）赋予更高采样权重，并引入 1 秒间隔的 k 帧历史图像与机器人动作作为上下文。使用 DeepSeek-R1 reasoning LLM 生成自然语言指令（任务名称、历史/当前/下一子任务），大幅提升指令多样性与语义覆盖。G0-VLM is instruction-tuned from the open-source Qwen2.5-VL, combining the human-annotated subtasks in the Galaxea data with synthesized high-level instructions. During training, key frames (subtask termination or gripper state change) are given a higher sampling weight, and k history frames at one-second intervals together with the robot actions are introduced as context. The DeepSeek-R1 reasoning LLM is used to generate natural-language instructions (task name, previous/current/next subtask), greatly increasing instruction diversity and semantic coverage.

03 实验Experiments

实验设计围绕核心问题：预训练数据如何影响 VLA？ 评测指标为 "progress score"（每任务 10 次测试运行的平均分），基准任务涵盖桌面操作、少样本迁移与移动操作。The experiments are organized around a core question: how does pre-training data affect a VLA? The evaluation metric is the "progress score" (the average over 10 test runs per task), and the benchmark tasks cover tabletop manipulation, few-shot transfer and mobile manipulation.

评测基准任务Evaluation Benchmark Tasks

**Figure 8**：四个评测基准任务。**Table Bussing**（满分 6 分）：整理杂乱桌面（钢笔、耳机、书籍）；**Microwave Operation**（满分 5 分）：将食物放入微波炉；**Bed Making**（满分 4 分）：整理凌乱被褥（需全身协调）；**Blocks Stacking**（满分 6 分）：用积木拼出指定词语。**Figure 8**: the four evaluation benchmark tasks. **Table Bussing** (max score 6): tidy a cluttered table (pens, headphones, books); **Microwave Operation** (max score 5): put food into the microwave; **Bed Making** (max score 4): tidy a messy bed (requires whole-body coordination); **Blocks Stacking** (max score 6): spell a given word with blocks.

G0-VLA 预训练权重对比Comparison of G0-VLA Pre-training Weights

在每任务 100 条训练轨迹的微调设置下，对比以下配置：Under a fine-tuning setting with 100 training trajectories per task, the following configurations are compared:

G0 (Full)：Stage-1 + Stage-2 预训练（400 小时）G0 (Full): Stage-1 + Stage-2 pre-training (400 hours)
G0 (Stage-2 400h)：仅 Stage-2（400 小时）G0 (Stage-2 400h): Stage-2 only (400 hours)
G0 (Stage-2 200h)：仅 Stage-2（200 小时）G0 (Stage-2 200h): Stage-2 only (200 hours)
G0 (Stage-1)：仅 Stage-1 预训练G0 (Stage-1): Stage-1 pre-training only
G0 (Scratch)：无动作预训练G0 (Scratch): no action pre-training
π₀：官方发布的预训练权重π₀: the officially released pre-trained weights

**Figure 9**：不同预训练 VLA 的微调基准结果。G0 (Full) 在平均 progress score 上表现最优，在 Table Bussing、Microwave Operation、Bed Making 等抓取任务中领先；G0 (Stage-2) 在语言跟随、动作一致性与全身控制上表现最佳；G0 (Stage-1) 是所有预训练模型中最差的，"underscoring the importance of single-embodiment pre-training"。**Figure 9**: fine-tuning benchmark results for differently pre-trained VLAs. G0 (Full) attains the best average progress score and leads on grasping tasks such as Table Bussing, Microwave Operation and Bed Making; G0 (Stage-2) performs best on language following, action consistency and whole-body control; G0 (Stage-1) is the weakest of all pre-trained models, "underscoring the importance of single-embodiment pre-training".

关键发现：单体具身预训练 vs. 跨具身预训练Key Finding: Single-Embodiment vs. Cross-Embodiment Pre-training

少样本迁移实验（每任务仅 20 条轨迹，10 个 epoch）表明：含 Stage-2 的模型显著优于无 Stage-2 的模型，动作更流畅稳定；而"Stage-1 alone do not show a clear advantage over models trained from scratch"，说明跨具身动作预训练单独使用可能不足。Bed Making（全身协调任务）的 per-skill progress 分析进一步印证：Stage-2 单体预训练大幅改善底盘与躯干控制，而跨具身预训练（Stage-1、π₀）"yields weaker performance, in some cases worse than training from scratch"，说明具身鸿沟在全身动作控制上尤为突出。Few-shot transfer experiments (only 20 trajectories per task, 10 epochs) show that models including Stage-2 clearly outperform those without it, with smoother and more stable actions; whereas "Stage-1 alone do not show a clear advantage over models trained from scratch", suggesting that cross-embodiment action pre-training on its own may be insufficient. The per-skill progress analysis on Bed Making (a whole-body coordination task) corroborates this: Stage-2 single-embodiment pre-training greatly improves base and torso control, while cross-embodiment pre-training (Stage-1, π₀) "yields weaker performance, in some cases worse than training from scratch", showing that the embodiment gap is especially pronounced for whole-body motion control.

G0-VLM 指令准确率（%）G0-VLM Instruction Accuracy (%)

模型Model	Table Bussing	Microwave Operation	Make the Bed	Build Blocks
Gemini-2.5-pro	32.0	15.8	54.2	55.0
Qwen2.5-VL-72B	26.3	16.8	48.1	21.7
Qwen2.5-VL-32B	21.3	14.8	54.2	21.0
Qwen2.5-VL-7B	26.3	17.2	46.9	24.7
G0-VLM（本文）G0-VLM (ours)	83.3	74.2	78.2	75.6

G0-VLM "surpasses baseline accuracy by over 50%"，验证了机器人应用需要精确对齐的动作原语，而非仅仅通用视觉-语言理解能力。G0-VLM "surpasses baseline accuracy by over 50%", confirming that robotic applications need precisely aligned action primitives rather than merely general vision-language understanding.

04 局限性Limitations

Note：论文未设独立的 Limitations 节。以下条目均为从设计与实验现象中推断（inferred）的局限，非作者明确陈述。Note: the paper has no standalone Limitations section. All items below are limitations inferred from its design and experimental observations, not explicitly stated by the authors.

单一硬件平台的泛化约束Generalization constraints of a single hardware platform

整个数据集以 Galaxea R1 Lite 单一机器人平台采集，以确保动作空间一致性。这是 Stage-2 单体预训练有效的关键前提，但也意味着模型在其他具身形态（不同 DoF、不同传感器配置）上的直接泛化受限——新平台需重新采集数据并经历相应的训练阶段。（inferred）The whole dataset is collected on the single Galaxea R1 Lite robot platform in order to guarantee action-space consistency. This is the key precondition for Stage-2 single-embodiment pre-training to be effective, but it also means that direct generalization to other embodiments (different DoF, different sensor configurations) is limited — a new platform would require re-collecting data and going through the corresponding training stages. (inferred)

跨具身预训练（Stage-1）的负迁移风险Negative-transfer risk of cross-embodiment pre-training (Stage-1)

实验显示，对于具身鸿沟较大的任务（如全身协调的 Bed Making），Stage-1 的跨具身预训练"in some cases worse than training from scratch"，说明动作空间不对齐时，大规模跨具身数据可能带来负迁移。如何在跨具身广度与单体一致性之间取得平衡，仍是开放问题。（inferred）Experiments show that for tasks with a large embodiment gap (such as the whole-body coordination task Bed Making), the cross-embodiment pre-training of Stage-1 is "in some cases worse than training from scratch", indicating that when action spaces are misaligned, large-scale cross-embodiment data may bring negative transfer. How to balance cross-embodiment breadth against single-embodiment consistency remains an open question. (inferred)

评测场景与真实长尾分布的差距Gap between the evaluation scenes and the real-world long tail

当前基准任务（Table Bussing、Microwave Operation、Bed Making、Blocks Stacking）设计合理但数量有限，每任务最多 100 条训练轨迹。论文中对移动操作（mobile manipulation）等更复杂场景的定量评测较少，泛化能力在更广泛的长尾任务上尚待验证。（inferred）The current benchmark tasks (Table Bussing, Microwave Operation, Bed Making, Blocks Stacking) are sensibly designed but limited in number, with at most 100 training trajectories per task. The paper reports relatively little quantitative evaluation of more complex settings such as mobile manipulation, so generalization over a broader long tail of tasks remains to be verified. (inferred)

双系统异步通信的延迟与协调开销Latency and coordination overhead of asynchronous dual-system communication

G0-VLM（慢系统）与 G0-VLA（快系统）异步运行；慢系统的规划延迟可能在高动态任务中影响实时性。论文未详细报告两系统间通信延迟对任务成功率的影响。（inferred）G0-VLM (the slow system) and G0-VLA (the fast system) run asynchronously; the planning latency of the slow system may affect real-time behaviour in highly dynamic tasks. The paper does not report in detail how the communication latency between the two systems affects task success rate. (inferred)