$ \pi_{0} $: A Vision-Language-Action Flow Model for General Robot Control

Physical Intelligence

目录 / Outline

$ \pi_{0} $: A Vision-Language-Action Flow Model for General Robot Control p.1
I. INTRODUCTION p.2
II. RELATED WORK p.3
III. OVERVIEW p.4
IV. THE $ \pi_{0} $ MODEL p.4
V. DATA COLLECTION AND TRAINING RECIPE p.5
A. Pre-training and post-training p.5
B. Language and high-level policies p.6
C. Robot system details p.6
VI. EXPERIMENTAL EVALUATION p.6
A. Evaluating the base model p.7
B. Following language commands p.8
C. Learning new dexterous tasks p.9
D. Mastering complex multi-stage tasks p.10
VII. DISCUSSION, LIMITATIONS, AND FUTURE WORK p.11
ACKNOWLEDGEMENTS p.12

引言

机器人学习在实现通用、灵活和灵巧的机器人系统方面具有巨大潜力，但面临数据稀缺、泛化性和鲁棒性等挑战。开发能够执行各种任务的通用机器人策略（即机器人基础模型）是解决这些挑战的关键，但需要克服规模、模型架构和训练配方三大瓶颈。本文提出$\pi_0$模型，一种基于预训练视觉-语言模型（VLM）的流匹配架构。通过在7种机器人配置和68个任务的多样跨具身数据集上进行预训练，并结合预训练/后训练配方，$\pi_0$能够实现零样本控制、遵循语言指令以及通过微调学习新技能，在洗衣折叠、桌面清理和箱体组装等灵巧任务上表现出色。

方法

如何整合多样化的数据源并利用互联网规模的语义知识？ → 使用预训练视觉-语言模型（VLM）作为骨干网络，并通过跨具身训练结合来自多种机器人的数据。

如何建模高频、灵巧的机器人动作？ → 采用动作分块（action chunking）架构，使用流匹配（flow matching）生成连续动作分布。

如何将流匹配集成到VLM中？ → 设计单独的动作专家（action expert）模块，该模块具有独立的权重，仅处理机器人特定的状态和动作令牌，并与VLM骨干并行工作。

如何训练出既鲁棒又高效的策略？ → 采用两阶段训练配方：先在多样化的大规模数据上预训练以获得广泛能力和恢复行为，再在高质量、任务特定的数据上进行后训练以掌握高效流畅的执行策略。

结果

归一化任务得分（零样本）：$\pi_0$在所有零样本任务上表现最佳，在衬衫折叠和简单收拾任务上接近完美，大幅超越所有基线模型。

对比方法	数值	引用
OpenVLA (全数据训练)	明显低于$\pi_0$	[24]
Octo (全数据训练)	低于$\pi_0$	[50]
OpenVLA (仅UR5e数据)	高于全数据OpenVLA但仍远低于$\pi_0$	[24]
$\pi_0$-small	低于$\pi_0$，但高于OpenVLA和Octo
$\pi_0$ (计算量匹配，160k步)	略低于完整$\pi_0$，仍优于所有基线

§VI-Ap.8The results, shown in Figure 7, show that \pi_0 attains by far the best results across the board on all the zero-shot tasks, with near perfect success rates on shirt folding and the easier bussing tasks, and large improvements over all baselines.

Fig. 6（第 7 页）：Zero-shot evaluation tasks: To evaluate our base model, we run it after pre-training on five tasks: shirt folding, bussing easy, bussing hard, grocery bagging, and toast out of toaster. The tasks require a combination of dexterous manipulation, multi-stage behaviors, and semantic recognition.

Fig. 7（第 7 页）：Zero-shot evaluation results: We evaluate $ \pi_0 $ trained for the full 700k steps, a version trained for 160k steps that matches the number of updates for baseline models, $ \pi_0 $-small, and three baselines: OpenVLA and Octo trained on all of our data, and OpenVLA trained only on the UR5e tasks (which we found to work a bit better on these tasks). Across all tasks and all comparisons, even the “parity” version of our model outperforms all baselines, and the full version of our model achieves the best results by a large margin.

语言指令遵循准确率：$\pi_0$的语言遵循准确率显著优于$\pi_0$-small，这转化为在专家人类指导（—human）和高层策略指导（—HL）下更好的任务表现。

对比方法	数值	引用
$\pi_0$-small (flat)	显著低于$\pi_0$
$\pi_0$-small (human)	低于$\pi_0$-human
$\pi_0$-small (HL)	未报告（可能更低）

§VI-Bp.9The results in Figure 9, averaging over 10 trials per task, show that the language following accuracy of \pi_0 is significantly better than that of \pi_0-small.

Fig. 8（第 8 页）：The tasks in our language evaluation. We evaluate our model on 3 different language-conditioned tasks, each of which requires following a sequence of intermediate language commands. The tasks involve bussing a table (top) to put dishes in a bin and garbage in a trash bin, setting a table (middle) by taking items out of a bin, and packing a shopping bag (bottom).

Fig. 9（第 8 页）：Language evaluation. We compare “flat” versions of our policies, $ -flat $, which receive only the overall task command (e.g., “bag the groceries”) with a method that receives intermediate commands from a human expert, —human, or a high-level VLM policy, —HL. We also compare our model to a small non-VLM variant under the “expert” condition, $ \pi_0 $ and $ \pi_0 $-small, in terms of language following accuracy. The results show a significant improvement with $ \pi_0 $ from intermediate language commands provided by a human expert and to a lesser degree by an autonomous high-level policy. Crucially, due to $ \pi_0 $-small’s language following performance, it does not gain with the addition of a high-level expert.

微调后任务得分（不同数据量）：$\pi_0$在多数任务上优于其他方法，预训练带来的改进在数据量较小时尤其明显，有时提升可达2倍。

对比方法	数值	引用
OpenVLA (微调)	显著低于$\pi_0$	[24]
Octo (微调)	显著低于$\pi_0$	[50]
ACT (从零训练)	在部分任务上接近但整体不如$\pi_0$	[57]
Diffusion Policy (从零训练)	在部分任务上接近但整体不如$\pi_0$	[9]
$\pi_0$ (从零训练)	低于预训练$\pi_0$

§VI-Cp.10The results show that \pi_0 generally outperforms other methods … the pre-trained model is frequently better than the non-pre-trained model, sometimes by as much as 2x.

Fig. 10（第 9 页）：Fine-tuning evaluation tasks: We fine-tune our model to a variety of downstream tasks that are distinct from the tasks seen in pre-training. Our tasks represent a range of similarity from the pre-training tasks, with tasks that are most similar to pre-training (stack bowls and towel folding), a task that introduces an unseen new element (a microwave), and tasks that require new motions and new object types (Franka items in drawer and paper towel replacement).

Fig. 11（第 10 页）：Fine-tuning with varying amounts of data. $\pi_{0}$ can learn some easier tasks even with smaller amounts of data, and the pre-trained model often attains a larger improvement over the model trained from scratch.

复杂多阶段任务平均得分（10次试验）：完整的预训练$\pi_0$模型在所有任务上均获得超过最高分50%的成绩，且通常优于消融实验模型，在最具挑战性的任务上改进尤为显著。

对比方法	数值	引用
消融模型（无预训练或小型模型）	低于完整$\pi_0$

§VI-Dp.11The full pre-trained \pi_0 model attains more than 50% of the maximum score across all of the tasks, and typically outperforms the ablations, with especially significant improvements on the hardest tasks.

Fig. 2（第 2 页）：$\pi_{0}$ controls a mobile manipulator to fold laundry. Our model is pre-trained on diverse data from 7 distinct robot configurations and 68 tasks, and can then be used in zero-shot or fine-tuned to complex downstream tasks, as in the case of this laundry folding policy, which fetches laundry from a dryer, packs it into a hamper, brings the hamper to a folding table, and then folds each article of clothing.

Fig. 12（第 10 页）：We evaluate a range of complex and temporally extended tasks. This includes: folding laundry from a bin with a stationary (a) or mobile (b) robot, bussing a real lunch table (c), assembling a box (d), packing eggs into a carton (e), and packing food into a to-go box (f). These tasks require combining dozens of individual behaviors, such as grasping, stacking, folding, and flattening, generalization to a huge variety of object configurations, and complex physical properties, such as deformable objects or flexible cardboard.

Fig. 13（第 11 页）：Post-training results on complex tasks in terms of average scores over 10 trials. The full pre-trained $ \pi_{0} $ model attains more than 50% of the maximum score across all of the tasks, and typically outperforms the ablations, with especially significant improvements on the hardest tasks.

实现

模型基于PaliGemma视觉-语言模型，通过添加动作专家模块将VLM转换为VLA。输入包括多幅RGB图像、语言指令和机器人本体状态，经视觉编码器和文本编码器处理后，由动作专家通过流匹配生成长度为50的动作块。训练分两阶段：预训练使用包含OXE和自采数据的混合集（10,000+小时），后训练使用高质量任务特定数据微调。

**Fig. 1**（第 1 页）：Our generalist robot policy uses a pre-trained vision-language model (VLM) backbone, as well as a diverse cross-embodiment dataset with a variety of dexterous manipulation tasks. The model is adapted to robot control by adding a separate action expert that produces continuous actions via flow matching, enabling precise and fluent manipulation skills. The model can then be prompted for zero-shot control or fine-tuned on high-quality data to enable complex multi-stage tasks, such as folding multiple articles of laundry or assembling a box.

**Fig. 3**（第 4 页）：Overview of our framework. We start with a pre-training mixture, which consists of both our own dexterous manipulation datasets and open-source data. We use this mixture to train our flow matching VLA model, which consists of a larger VLM backbone and a smaller action expert for processing robot states and actions. The VLM backbone weights are initialized from PaliGemma [5], providing representations learned from large-scale Internet pre-training. The resulting $ \pi_{0} $ model can be used to control multiple robot embodiments with differing action spaces to accomplish a wide variety of tasks.

VLM骨干网络 — 处理图像和语言输入，提取语义特征。
设计：使用PaliGemma 3B模型，采用晚期融合架构，将图像嵌入投影到与文本令牌相同的嵌入空间。
原理：继承来自大规模互联网预训练的视觉和语言表示。

动作专家 — 处理机器人状态和动作令牌，生成动作块。
设计：独立的Transformer权重（300M参数），全双向注意力掩码，与VLM骨干共享输入但具有独立的前向网络。
原理：通过处理噪声动作令牌和观测来预测流匹配的向量场，从而建模连续动作分布。

流匹配训练与推理 — 训练时通过流匹配损失监督动作令牌；推理时从噪声逐步去噪生成动作块。
设计：采用条件流匹配，线性高斯概率路径$q(\mathbf{A}_t^\tau|\mathbf{A}_t) = \mathcal{N}(\tau\mathbf{A}_t, (1-\tau)\mathbf{I})$，网络输出向量场$\mathbf{v}_\theta$；推理使用10步前向欧拉积分。
原理：流匹配通过训练一个向量场来将一个简单分布转换为目标动作分布。训练损失：$\mathcal{L} = \mathbb{E}_{t,\epsilon}[\| \mathbf{v}_\theta(\mathbf{A}_t^\tau, \mathbf{o}_t) - (\epsilon - \mathbf{A}_t) \|^2]$，推理从$\tau=0$到$\tau=1$积分：$\mathbf{A}_t^{\tau+\delta} = \mathbf{A}_t^\tau + \delta \mathbf{v}_\theta(\mathbf{A}_t^\tau, \mathbf{o}_t)$。

$ \pi_{0} $: A Vision-Language-Action Flow Model for General Robot Control

引言

相关工作

方法

结果

实现