π₀: How a Robot Turns Understanding into Continuous Action

一份兼顾直觉、公式与路线思辨的双语读书笔记。配对版本：English 核心论文：π₀: A Vision-Language-Action Flow Model for General Robot Control 核验口径：arXiv v4，2026-01-08；历史措辞参考：arXiv v1，2024-10-31

这篇笔记使用五种内容标签：

论文事实：由 π₀ 论文直接报告，但不等于已有第三方独立复现；
作者声明：作者关于“首次、最大、SOTA”等判断，必须保留“据作者所知”等限定；
代码快照：当前 openpi 公开实现呈现的状态，可能与论文时期接口不同；
通俗解释：为了帮助理解而做的转述，不是论文原句；
我的思考：基于证据形成的路线判断，不冒充论文结论。

一、先用一分钟抓住 π₀

1. π₀ 是什么

π₀（读作 pi-zero）是 Physical Intelligence 发布的第一代通用机器人策略。它不是“Physical Intelligence Zero”的正式全称，也不是一个显式预测未来视频的 World Model。它是一种 Vision-Language-Action policy：根据视觉、语言指令和机器人本体状态，直接生成连续动作。

最简洁的架构表达是：

\pi_0 = \underbrace{\text{VLM 语义骨干}}_{\text{编码看到了什么、任务要求什么}} + \underbrace{\text{连续 Flow 动作专家}}_{\text{决定接下来怎样运动}}

但 Flow Matching 只是动作生成机制，不是 π₀ 的全部。完整配方还包括：

\boxed{ \pi_0 = \text{VLM 语义} + \text{连续动作专家} + \text{跨机器人数据} + \text{pretraining/post-training} + \text{滚动执行} }

论文事实：

视觉语言骨干采用约 30 亿参数的 PaliGemma；
action expert 约 3 亿参数，总参数量约 33 亿；
模型一次生成 $H=50$ 个连续物理动作；
推理时对同一个动作块执行 $K=10$ 次 Euler 更新。

2. 它弥合了什么设计缺口

π₀ 出现之前，两类模型各自解决了问题的一部分：

技术路线	擅长什么	主要缺口
VLM / 早期 VLA	理解图像、语言、物体和开放词汇指令	离散、自回归动作输出不天然适合高频精细控制
Diffusion / 连续控制策略	生成平滑、多模态的连续动作	缺少大型 VLM 的语义知识和语言迁移能力
π₀	将大型 VLM 与连续动作生成器接在一起	仍受机器人示范数据覆盖范围约束

π₀ 的关键选择是：不要求一个语言模型同时勉强承担语义理解和底层马达控制，而是让 VLM 提供任务相关的视觉语言上下文，让 action expert 生成连续动作。这一“语义骨干—动作专家”接口，是论文最重要且最耐久的设计。

二、机器人实际执行时发生了什么

假设指令是：“把桌上的盘子和杯子收进周转箱。”

1. 观察现实

机器人接收：

桌面相机、腕部相机等 2–3 路 RGB 图像；
自然语言指令；
当前关节角、夹爪状态等本体状态。

2. 形成任务上下文

PaliGemma 编码与任务有关的视觉语言信息：哪些物体是盘子、杯子和箱子，当前目标是什么，物体与指令有什么关系。这里不是“VLM 先说出一句话，再交给动作解码器”；action expert 通过 attention 读取内部隐藏表示。

3. 从噪声生成动作

action expert 从一块高斯噪声开始，经过 10 次 Flow 更新，得到一段包含 50 个连续动作的 action chunk。教学上可以把它描述为：

靠近盘子 → 调整手腕 → 闭合夹爪 → 抬起 → 移向箱子 → 放下。

真实输出是连续数值，而不是自然语言步骤。

4. 只执行动作块前缀

π₀ 虽然预测 50 步，却不会盲目执行全部动作：

20 Hz 平台执行前 16 步，约 0.8 秒后重新规划；
50 Hz 平台执行前 25 步，约 0.5 秒后重新规划。

5. 重新观察并滚动规划

机器人重新拍摄真实场景。如果物体移动、夹持失败或姿态偏离，下一轮动作会根据新观察重新生成：

观察现实
→ 形成任务上下文
→ 生成 50 步动作块
→ 执行其中一部分
→ 重新观察现实
→ 再生成

**关键边界：**π₀ 在行动之后重新观察真实世界；它没有在行动之前显式生成一段未来视频来预测世界变化。

三、模型的输入和输出

1. 当前观察

物理时间 $t$ 的观察可写为：

o_t=[I_t^1,\ldots,I_t^n,\ell_t,q_t]

其中：

$I_t^i$ ：第 $i$ 路相机图像；
$\ell_t$ ：语言指令；
$q_t$ ：机器人本体状态，例如关节角和夹爪状态。

$q_t$ 不是模型内部状态，也不是所谓“action 状态”，而是机器人当前真实的物理状态。

2. 单步动作向量

a_t\in\mathbb R^d

$a_t$ 表示一个物理时间步的连续控制向量； $d$ 是单步动作维数，不同机器人可以不同。

3. 动作块

A_t=[a_t,a_{t+1},\ldots,a_{t+H-1}] \in\mathbb R^{H\times d}

π₀ 使用 $H=50$ 。若展平：

\operatorname{vec}(A_t)\in\mathbb R^D, \qquad D=Hd

几个容易混淆的概念必须分开：

概念	表示	含义
action vector	$a_t$	一个物理时间步的动作
action chunk	$A_t$	连续 $H$ 个物理动作
action slot	Transformer 序列位置	承载一个连续动作向量的内部槽位
discrete action token	词表 ID	某些 VLA 使用的离散动作表示

π₀ 的 action slot 承载连续向量，不等于语言词表中的离散 token。

4. 模型真正学习的对象

\boxed{ p_{\mathrm{data}}(A_t\mid o_t) }

通俗地说：给定当前画面、语言目标和机器人姿态，生成一段合理的未来动作。生成的是动作，不是未来图像或未来世界状态。

四、架构：VLM 与 action expert 怎样协作

1. 两套专长

π₀ 可以理解为同一系统中的两套专长权重：

PaliGemma VLM，约 3B 参数：处理图像和语言，提供物体、场景和指令语义；
action expert，约 300M 参数：处理机器人状态、带噪动作和 Flow 时间，预测连续动作的修改方向。

一个粗略比喻是“领班 + 编舞师”：领班掌握现场与目标，编舞师把任务意图转成各关节的协调运动。这个比喻只描述分工；模型内部交换的是隐藏表示，并非可读语言。

2. 它不是普通的稀疏 MoE

图像和语言固定进入 VLM 权重，本体状态和动作固定进入 action-expert 权重，两者通过 self-attention 交换信息。系统没有学习一个路由器，临时决定每个 token 应该进入哪个专家。

更准确地说，π₀ 使用固定的模态分工，让语义骨干与动作专家在同一 Transformer 式注意力系统中协作。

3. 为什么不把动作直接当作语言 token

机器人动作天然是连续数值，并要求多个关节在时间上协调。把动作量化成离散词表并逐 token 输出，可能引入：

量化误差；
自回归延迟；
多关节和多时间步之间的协调困难。

π₀ 一次联合生成整个连续动作块，使 50 个动作位置能够相互协调。

五、Conditional Flow Matching：从噪声学习到动作的流

1. 为什么从噪声开始

高斯噪声是容易采样的简单分布：

\epsilon\sim\mathcal N(0,I)

$\epsilon$ 与动作块 $A_t$ 形状相同。这里的噪声不是注入真实机器人的物理干扰，也不是普通正则化噪声；它是生成模型的起始分布。

模型要学习的是：

\text{简单高斯分布} \longrightarrow \text{给定观察条件后的动作分布}

2. 构造训练路径

采样 Flow 时间：

\tau\in[0,1]

在噪声和真实示范动作之间做线性插值：

\boxed{ A_t^\tau=(1-\tau)\epsilon+\tau A_t }

端点为：

A_t^0=\epsilon, \qquad A_t^1=A_t

$\tau=0$ 是纯噪声， $\tau=1$ 是真实动作块，中间状态是一块“半噪声、半动作”的候选动作。这条直线位于生成空间，不代表机械臂在物理空间中沿直线移动。

3. 得到配对训练目标

将插值式写成：

A_t^\tau=\epsilon+\tau(A_t-\epsilon)

对 $\tau$ 求导：

\boxed{ \frac{dA_t^\tau}{d\tau}=A_t-\epsilon }

因此，对每一组采样的 $(\epsilon,A_t)$ ，监督目标速度是：

u=A_t-\epsilon

通俗地说，它告诉模型这块尚未完成的候选动作应该朝什么方向、以多大幅度修改。

4. action expert 学习条件向量场

模型预测：

v_\theta(A_t^\tau,o_t,\tau)

训练损失为：

\boxed{ \mathcal L(\theta) = \mathbb E \left[ \left\| v_\theta(A_t^\tau,o_t,\tau) -(A_t-\epsilon) \right\|_2^2 \right] }

5. 必须保留的专业限定

训练时，每个噪声—示范动作配对都有目标 $A_t-\epsilon$ ；实际推理时，模型并不知道某个预先指定的真实动作 $A_t$ 。

在理想的无限数据和均方误差优化下，模型学习的是条件向量场：

v^*(x,o,\tau) = \mathbb E \left[ A_t-\epsilon \mid A_t^\tau=x,\ o_t=o,\ \tau \right]

这意味着 π₀ 学习的是怎样把高斯噪声分布运输成当前条件下的动作分布，而不是检索某一条训练示范。不同初始噪声仍可对应不同的合理动作；条件期望向量场不等于简单输出一条“平均动作”。

六、推理：10 次 Euler 更新究竟做了什么

1. 从新噪声开始

\hat A_t^{(0)}\sim\mathcal N(0,I)

推理时有当前观察 $o_t$ ，但没有真实动作答案 $A_t$ 。

2. 学到的 ODE

\frac{d\hat A_t^\tau}{d\tau} = v_\theta(\hat A_t^\tau,o_t,\tau)

3. 使用 Forward Euler 离散求解

π₀ 设置：

K=10, \qquad \delta=\frac{1}{K}=0.1

更新公式：

\boxed{ \hat A_t^{(k+1)} = \hat A_t^{(k)} + \frac{1}{K} v_\theta \left( \hat A_t^{(k)},o_t,\frac{k}{K} \right) }

其中 $k=0,1,\ldots,9$ ，最后得到：

\hat A_t=\hat A_t^{(10)}

4. Euler 法没有证明“10 步必然得到真实动作”

在一个教学特例中，若假设速度始终为已知常数：

v_\theta=A_t-\epsilon

则：

\hat A_t^{(k)} = \epsilon+\frac{k}{K}(A_t-\epsilon)

当 $k=K$ 时，确实有：

\hat A_t^{(K)} = \epsilon+(A_t-\epsilon) =A_t

这只验证了一条已知、恒速的配对直线路径，不能证明真实模型中 10 步理论上必需、一定足够，或每个噪声样本都会逼近某个指定示范动作。事实上，如果速度恒定且终点已知，一步 $\delta=1$ 也能到达终点。

真实模型需要多步，因为学到的向量场随候选动作位置、当前观察和 $\tau$ 改变，同时存在网络近似误差与 Euler 离散误差。因此，10 步是生成质量、数值精度和推理计算之间的工程折中，不是“十步收敛定理”。

5. $H=50$ 与 $K=10$ 完全不同

\boxed{ 10\text{ 次 Flow 更新} \longrightarrow 1\text{ 个包含 50 步物理动作的 action chunk} }

$H=50$ ：动作块包含的物理时间步数；
$K=10$ ：生成同一个动作块时的数值积分步数。

七、训练与运行不要混在一起

训练阶段

真实观察 o_t 与真实动作块 A_t
→ 采样噪声 ε
→ 采样 Flow 时间 τ
→ 构造中间动作 A_t^τ
→ 监督 action expert 预测 A_t - ε
→ 更新模型参数

运行阶段

当前观察 o_t 与一块新噪声
→ 做 10 次 Euler 更新
→ 得到 50 步动作块
→ 执行动作块前缀
→ 重新观察真实世界
→ 滚动重规划

训练时存在真实动作监督；运行时没有真实答案，只能沿学到的条件向量场采样。

八、公式之外：数据与训练配方同样是核心

1. 数据规模

论文报告：

超过 10,000 小时机器人操作数据；
约 903M 自有 timesteps；
7 类 robot configurations；
68 个宽任务；
开放数据占训练采样混合的约 9.1%；
不同机器人的状态与动作统一填充到最大 18 维接口。

903M 指时间步，不是 903M 条完整轨迹；9.1% 是采样混合占比，不是原始时间步占比。

2. 广泛预训练

预训练数据覆盖多机器人、多任务、多物体和多场景，也包含不完美动作、偏离状态与恢复过程。它的目标不是让每项任务立即达到最高熟练度，而是扩大模型见过的状态和行为范围。

3. 高质量 post-training

任务后训练使用更一致、更熟练、更有针对性的示范，使动作趋于稳定和流畅。简单任务可能需要约 5 小时专项数据，复杂任务可能需要 100 小时以上。

π₀ 没有消灭任务数据收集，而是把专项数据的作用从“从零学习全部能力”转变为“校准和精修已有能力”。

4. 整套配方

\boxed{ \text{广泛预训练负责能力覆盖} \quad+\quad \text{高质量后训练负责动作熟练度} }

这在方法论上类似“先广泛预训练，再有针对性地适配”，但 π₀ 的任务后训练不应简单等同于语言模型的 RLHF 或 alignment。

九、实验究竟证明了什么

论文从四个层级验证系统：

证据层级	主要结果	必须保留的限定
Direct prompting	五个任务中 π₀ 的归一化进度分约为 0.75–1.00，明显高于论文基线	这些任务族存在于预训练中，不是严格未见任务
语言跟随	能利用人类或高层 VLM 给出的中间语言指令	π₀-small 同时改变规模、初始化和架构，不是干净的单因素消融
新任务适配	使用 1/5/10 小时数据微调时，预训练通常提高样本效率	并非每个任务、每个数据点都获胜
复杂任务	展示洗衣、收桌、装盒等 5–20 分钟任务	“超过 50%”是部分进度分，不等于完整成功率

其中 0.75–1.00 是对论文图 7 的近似读图范围，并非论文表格给出的精确数值。

1. v1 与 v4 的 zero-shot 口径

首发 v1 使用了 zero-shot，但同一版本说明五个基础评测任务族存在于预训练中。当前 v4 已改为 direct prompting / out-of-box。

准确的解释是：模型没有针对相应测试版本做任务专门 post-training，不等于它从未见过相关任务族、机器人或行为分布。

2. 长任务不等于完整自主规划

部分长任务依赖人类或独立高层 VLM 提供中间指令。π₀ 主要证明了通用底层策略能力，并没有由单个模型同时包办长期目标分解、持久记忆、成功验证、安全判断和底层连续控制。

十、π₀ 的核心创新究竟是什么

π₀ 不是下列任何单项概念的发明者：

RT-2 更早提出 VLA；
Diffusion Policy 更早探索连续生成式控制；
action chunking 并非由 π₀ 首创；
Octo 是更早的跨机器人通用策略；
OpenVLA 更早公开了基于 Internet 预训练 VLM 的 VLA。

π₀ 的贡献是系统整合：

用大型预训练 VLM 保留视觉语言语义；
用独立 action expert 处理本体状态与连续动作；
用 Flow Matching 联合生成高频 action chunk；
用万小时跨机器人数据进行基础预训练；
用高质量 post-training 把广泛能力变成熟练行为；
在柔性物体、双臂协调和长任务上做大规模真实机器人展示。

论文将贡献定位为整合型创新，只作了带 “to our knowledge” 限定的首创声明：据作者所知，它是首个用于灵巧控制、生成高频动作块的 flow-matching VLA。

因此，更稳妥的历史评价是：

π₀ 是 Physical Intelligence 的 π 系列和连续 action-expert VLA 路线的定义性起点，但不是整个 VLA 或具身智能领域的第一篇工作。

十一、我的思考：π₀ 与 WM/WAM 路线有什么区别

本章属于路线分析，不是 π₀ 论文已经证明的结论。WAM 是仍在形成中的非标准化术语，不同研究对其边界和耦合方式并没有统一定义。

1. 根本区别不是“语言介质 vs 视频介质”

我原来的直觉是：π₀ 与 LLM/VLM 的关联更深，VLA 的信息更像语言，而 WM/WAM 的信息更像视频或 latent。这个直觉抓住了信息侧重点，却把中间表示说得过于简单。

更准确的区分是：

π₀ 的语义骨干来自 PaliGemma，因此重视物体、指令和任务语义；
VLM 交给 action expert 的不是一段可读语言，而是图像、语言和状态形成的隐藏上下文；
WM/WAM 路线让未来世界的时空结构或预测监督实质参与动作学习。

π₀ 式直接策略学习：

p(A_t\mid o_t)

它问：“根据当前观察，我现在应该怎样行动？”

World Model 可以学习：

p(z_{t+1:t+H}\mid z_t,A_t)

它问：“如果执行这些动作，未来世界可能怎样变化？”其中 $z$ 可以是图像、视觉 latent、状态或其他世界表征，不必是人类可读视频。

联合型 WAM 的一种教学抽象是：

p(A_t,z_{t+1:t+H}\mid o_t)

但这不是统一定义；具体系统也可以采用其他分解方式、训练期预测监督或部署期 action-only 输出。

2. “行动后看结果”与“行动前推演”是教学性对比

π₀ 的闭环是：

观察 → 直接行动 → 世界真实变化 → 再观察

World Model 路线可以采用：

观察 → 内部预测若干未来 → 比较后果 → 选择行动

这是一种帮助理解的典型对比，并不意味着所有 World Model 或 WAM 都会在部署时显式生成视频并搜索多个未来。

没有显式 World Model，也不等于 π₀ 毫无物理知识。为了从示范中生成有效动作，其参数可能编码与接触、物体和机器人动力学有关的行动规律。区别在于：π₀ 没有使用一个可单独检查的未来预测目标来训练这些规律，也不天然提供反事实模拟器。它可能“会做”，却不一定显式展示动作之后世界会怎样变化。

3. 决策路径可能更短，但不保证更快

直接策略不必先生成未来世界、评估候选轨迹再选择动作，因此可能拥有较短的决策路径。但：

π₀ 自己仍需 10 次 action-expert Flow 更新；
某些 WAM 只在训练时使用未来预测监督，部署时可以 action-only；
工程速度必须比较端到端 p50/p95 延迟、控制频率、硬件与重规划方式。

因此，π₀ 倾向于以更直接的动作接口换取执行效率；WM/WAM 倾向于以更丰富的动态表征换取后果推演能力。具体快慢需要实测。

4. 两条路线的真实权衡

维度	π₀ 式直接 VLA	WM/WAM-first 路线
主要学习对象	条件动作分布	世界转移、未来表征与动作的耦合
动作前主要信息	当前观察形成的 hidden context	当前观察加未来结构或预测监督
是否必须生成视频	否	也不一定，可以是 latent 或仅训练期监督
反事实推演	不显式提供	更容易支持“如果这样做会怎样”
控制接口	直接生成连续动作块	可联合生成，也可通过策略或 action-only 头输出
典型优势	语义接口清晰、执行链短、适合连续控制	动态信息丰富，适合规划和后果判断
典型风险	可能流畅地做错，却缺少显式成功验证	计算、内存、数据要求和模型误差可能更高

5. 长期更可能融合

我的判断是：World Model 更适合慢速推演、长期规划、后果判断和异常检测；VLA action expert 更适合快速、连续的底层执行。这是由两条路线的互补性推导出的架构判断，不是论文结论。

语言与目标理解
→ 世界动态推演
→ 任务与子目标规划
→ 连续动作生成
→ 真实世界反馈

因此，“VLA vs WAM”未来可能不是产品分类，而是同一机器人系统内部不同层次的能力。

十二、最大的优点与最耐久的缺点

1. 最大优点：建立“语义—运动接口”

π₀ 最耐久的贡献不是某个榜单分数，也不一定是 Flow Matching 永远最好，而是把大型 VLM 的通用语义能力与一个可以替换和扩展的连续动作专家接在一起。

未来可以更换 VLM 主干、动作编码、Flow 求解器、chunk 长度或机器人平台，但“通用语义骨干 + 连续控制专家”的分工仍可能保留。

2. 最大结构性缺点：能力边界受示范支持域约束

π₀ 本质上仍是离线示范驱动的条件行为克隆。它没有显式提供：

自己是否处于训练分布外的不确定性；
不理解时的拒绝执行；
动作之后任务是否成功的验证器；
持久世界状态与长期记忆；
通过在线交互持续学习的机制。

因此，它可能在陌生状态下流畅而自信地做错。更多数据能扩大覆盖面，但不会自动带来对未知状态的自知。

3. 论文证据的耐久限制：难以因果归因

核心万小时数据无法被第三方完整获得；
基线训练预算和动作接口不完全一致；
π₀ 与 π₀-small 同时改变参数量、初始化和架构；
多数条件约进行 10 次真实机器人试验，论文未报告置信区间；
复杂任务使用作者自建的部分进度量表。

openpi 现已公开代码和基础权重，因此“完全不开源”已经不准确；但原始 10,000 小时预训练仍不能被第三方完整复刻。

论文有力证明的是整套系统配方在作者环境中有效，但没有干净分离 VLM、Flow、数据规模、post-training 与系统工程各自贡献了多少。

十三、最容易出现的误读

π₀ 是 Physical Intelligence Zero 的正式全称。 不是；直接称 π₀ 或 pi-zero。
π₀ 是 World Model。 不是；它不显式预测未来图像、状态或奖励。
VLM 先输出一句语言，再交给动作模型。 不是；action expert 读取内部隐藏上下文。
$H=50$ 表示动作是 50 维。 不是；它表示 50 个物理时间步。
10 次 Flow 更新会生成 10 个动作块。 不是；10 次更新共同生成一个动作块。
Forward Euler 证明 10 步必然到达真实动作。 不是；Euler 是数值求解器，10 步是工程选择。
50 Hz 表示模型每秒重新看图并完整推理 50 次。 不是；这是动作命令频率，系统约每 0.5–0.8 秒重规划。
Direct prompting 等于从未见过任务。 不是；基础任务族存在于预训练中。
长任务完全由 π₀ 单模型自主规划。 不完整；部分任务使用人类或高层 VLM 的中间指令。
实验已经证明 Flow Matching 是成功的唯一原因。 没有；论文主要证明整套 recipe 有效。

十四、一分钟复述

π₀ 是 Physical Intelligence 的第一代通用机器人策略。它用约 30 亿参数的 PaliGemma 编码图像和语言，再用约 3 亿参数的 action expert，通过 Conditional Flow Matching 生成包含 50 个连续物理动作的 action chunk。训练时，模型在高斯噪声和真实动作之间构造直线路径，学习候选动作应该怎样修改；推理时从新噪声开始，使用 10 次 Forward Euler 更新得到动作块，然后只执行其中一部分并重新观察现实。π₀ 真正的创新不是一条孤立公式，而是将 VLM 语义、连续动作专家、万小时跨机器人预训练和高质量 post-training 整合成一套 foundation-policy 配方。它与 World Model 的根本区别是：π₀ 直接学习条件动作分布，而 World Model 显式学习世界可能怎样变化。长期看，两条路线更可能在规划层和执行层融合。

十五、复盘自测

如果能够回答下面十个问题，就基本掌握了 π₀：

π₀ 的输入和输出分别是什么？
$a_t$ 、 $A_t$ 、 $H$ 、 $d$ 、 $D$ 分别表示什么？
action chunk 与 discrete action token 有什么区别？
为什么使用高斯噪声作为起点？
为什么配对训练目标是 $A_t-\epsilon$ ？
推理时为什么不存在一个已知的真实 $A_t$ ？
Forward Euler 在系统中负责什么，又不负责什么？
为什么 $H=50$ 与 $K=10$ 完全不同？
π₀ 的创新为什么是系统配方，而不是单一 Flow 公式？
π₀ 与 World Model/WAM 的学习目标有什么区别？

主要来源

核心论文：π₀ v4、π₀ v1
作者材料：PI π₀ 项目页、openpi
方法前史：PaliGemma 技术报告、Flow Matching、Rectified Flow
前置工作：RT-2、Diffusion Policy、Octo、OpenVLA
后续边界：π₀.5、Real-Time Chunking
WAM 路线参考：WAM Survey、VPP、DreamZero、Fast-WAM、GigaWorld-Policy
论文索引：Hugging Face paper page

版本与核验说明

公式、模型、数据与实验事实主要使用 π₀ arXiv v4；
“zero-shot”历史措辞使用 v1，并明确当前 v4 已改写；
论文推导使用 $\tau=0$ 为噪声、 $\tau=1$ 为动作；当前 openpi 代码采用相反的时间方向（ $t=1$ 为噪声、 $t=0$ 为动作， $dt<0$ ），两者通过 $t=1-\tau$ 等价，并不矛盾；
正文始终用一般动作维数 $d$ 。18 维是论文跨机器人数据接口的补齐维数；当前 openpi 的 Pi0Config 默认 action_dim=32，属于当前开源实现配置，不能混成同一个数字；
openpi 公共仓库状态于 2026-08-02 通过 GitHub connector 确认为公开仓库；
Hugging Face 页面沿用 2026-08-01 已核验快照，仅用于元数据与生态索引；
“我的思考”属于解释性推断，应与论文事实分开阅读。

A bilingual reading note that combines intuition, equations, and a critical comparison of technical approaches. Paired version: 中文 Core paper: π₀: A Vision-Language-Action Flow Model for General Robot Control Verification basis: arXiv v4, January 8, 2026; historical wording: arXiv v1, October 31, 2024

This note uses five content labels:

Evidence from the paper: reported directly by the π₀ paper, but not necessarily independently reproduced;
Author claim: judgments such as “first,” “largest,” or “state of the art,” with the authors’ original qualifications preserved;
Code snapshot: the state of the current public openpi implementation, which may differ from the paper-era interface;
Plain-language explanation: an explanatory restatement, not a quotation from the paper;
My interpretation: a route-level judgment derived from the evidence, not a claim made by the π₀ paper.

1. π₀ in One Minute

1.1 What is π₀?

π₀ (pronounced “pi-zero”) is Physical Intelligence’s first-generation generalist robot policy. “Physical Intelligence Zero” is not its formal name, and π₀ is not a world model that explicitly predicts future video. It is a vision-language-action (VLA) policy: given visual observations, a language instruction, and the robot’s proprioceptive state, it directly generates continuous actions.

The shortest architectural description is:

\pi_0 = \underbrace{\text{VLM semantic backbone}}_{\text{encodes what is observed and what the task requires}} + \underbrace{\text{continuous Flow action expert}}_{\text{determines how the robot should move next}}

Flow Matching, however, is only the action-generation mechanism. The full π₀ recipe also includes:

\boxed{ \pi_0 = \text{VLM semantics} + \text{continuous action expert} + \text{cross-embodiment data} + \text{pretraining/post-training} + \text{receding-horizon execution} }

Evidence from the paper:

The vision-language backbone is PaliGemma, with roughly 3 billion parameters.
The action expert has roughly 300 million parameters, bringing the total to about 3.3 billion.
The model generates $H=50$ continuous physical actions at a time.
At inference time, it applies $K=10$ Euler updates to the same action chunk.

1.2 What design gap does it bridge?

Before π₀, two model families had each solved a different part of the problem:

Approach	What it did well	Main gap
VLMs and early VLAs	Understanding images, language, objects, and open-vocabulary instructions	Discrete autoregressive action outputs were not naturally suited to high-frequency, fine-grained control
Diffusion and continuous-control policies	Generating smooth, multimodal continuous actions	Limited access to the semantic knowledge and language transfer of large VLMs
π₀	Connecting a large VLM to a continuous action generator	Remaining dependence on the coverage of robot demonstration data

π₀ makes a deliberate division of labor. It does not force one language model to perform both semantic interpretation and low-level motor control. The VLM supplies task-relevant visual-language context, while the action expert generates continuous motion. This semantic-backbone–action-expert interface is the paper’s most important and most durable design choice.

2. What Actually Happens When the Robot Acts

Suppose the instruction is: “Put the plates and cups on the table into the bus tub.”

2.1 Observe the real scene

The robot receives:

two or three RGB views, such as workspace and wrist-camera images;
a natural-language instruction;
its current proprioceptive state, including joint angles and gripper state.

2.2 Form task-relevant context

PaliGemma encodes task-relevant visual and linguistic information: which objects are plates, cups, and the tub; what goal is being requested; and how the objects relate to the instruction. The VLM does not first produce a readable sentence for an action decoder. Instead, the action expert attends to internal hidden representations.

2.3 Generate actions from noise

The action expert begins with a block of Gaussian noise. Ten Flow updates transform it into an action chunk containing 50 continuous actions. For teaching purposes, one might describe the sequence as:

approach the plate → adjust the wrist → close the gripper → lift → move toward the tub → release.

The actual output consists of continuous numerical values, not natural-language steps.

2.4 Execute only a prefix of the chunk

Although π₀ predicts 50 actions, it does not blindly execute all of them:

on a 20 Hz platform, it executes the first 16 actions and replans after roughly 0.8 seconds;
on a 50 Hz platform, it executes the first 25 actions and replans after roughly 0.5 seconds.

2.5 Reobserve and replan

The robot captures the real scene again. If an object moved, a grasp failed, or the pose drifted, the next action chunk is generated from the new observation:

observe the real scene
→ form task-relevant context
→ generate a 50-step action chunk
→ execute part of the chunk
→ observe the real scene again
→ generate again

Important boundary: π₀ observes the real world again after acting. It does not explicitly generate a future video before acting in order to predict how the world will change.

3. Model Inputs and Outputs

3.1 Current observation

The observation at physical time $t$ can be written as:

o_t=[I_t^1,\ldots,I_t^n,\ell_t,q_t]

where:

$I_t^i$ is the image from camera $i$ ;
$\ell_t$ is the language instruction;
$q_t$ is the robot’s proprioceptive state, such as its joint angles and gripper state.

$q_t$ is neither the model’s internal state nor an “action state.” It describes the robot’s actual physical state at that moment.

3.2 Single-step action vector

a_t\in\mathbb R^d

$a_t$ is the continuous control vector for one physical time step. The single-step action dimension $d$ can differ across robots.

3.3 Action chunk

A_t=[a_t,a_{t+1},\ldots,a_{t+H-1}] \in\mathbb R^{H\times d}

π₀ uses $H=50$ . When flattened:

\operatorname{vec}(A_t)\in\mathbb R^D, \qquad D=Hd

Four easily confused concepts must remain distinct:

Concept	Notation	Meaning
action vector	$a_t$	the action at one physical time step
action chunk	$A_t$	$H$ consecutive physical actions
action slot	a Transformer sequence position	an internal slot carrying one continuous action vector
discrete action token	a vocabulary ID	a discretized action representation used by some VLAs

An action slot in π₀ carries a continuous vector. It is not a discrete token from a language vocabulary.

3.4 What the model actually learns

\boxed{ p_{\mathrm{data}}(A_t\mid o_t) }

In plain language: given the current images, language goal, and robot pose, generate a plausible sequence of future actions. The output is a future action sequence, not a future image or future world state.

4. Architecture: How the VLM and Action Expert Work Together

4.1 Two specialized parameter sets

π₀ can be understood as two specializations within one system:

PaliGemma VLM, approximately 3B parameters: processes images and language and supplies object, scene, and instruction semantics;
action expert, approximately 300M parameters: processes proprioceptive state, noisy actions, and Flow time, then predicts how the continuous action should change.

As a rough analogy, the two components resemble a supervisor and a choreographer. The supervisor tracks the scene and objective; the choreographer converts that intent into coordinated joint motion. The analogy describes only the division of labor. Internally, the components exchange hidden representations rather than readable language.

4.2 It is not a conventional sparse MoE

Images and language are assigned to the VLM parameters, while proprioceptive state and actions are assigned to the action-expert parameters. The two exchange information through self-attention. There is no learned router that dynamically sends each token to a selected expert.

A more precise description is that π₀ uses fixed modality specialization, allowing a semantic backbone and an action expert to cooperate within a Transformer-style attention system.

4.3 Why not treat actions as language tokens?

Robot actions are inherently continuous and require coordination across joints and time. Quantizing them into a discrete vocabulary and emitting them token by token can introduce:

quantization error;
autoregressive latency;
difficulty coordinating multiple joints across multiple time steps.

π₀ instead generates the entire continuous action chunk jointly, allowing all 50 action positions to coordinate with one another.

5. Conditional Flow Matching: Learning a Flow from Noise to Action

5.1 Why start from noise?

Gaussian noise is a simple distribution from which samples are easy to draw:

\epsilon\sim\mathcal N(0,I)

$\epsilon$ has the same shape as the action chunk $A_t$ . This noise is neither a physical disturbance injected into the robot nor ordinary regularization noise. It is the generative model’s base distribution.

The model learns to transport samples from:

\text{a simple Gaussian distribution} \longrightarrow \text{the action distribution conditioned on the observation}

5.2 Construct the training path

First sample a Flow time:

\tau\in[0,1]

Then linearly interpolate between noise and a demonstrated action chunk:

\boxed{ A_t^\tau=(1-\tau)\epsilon+\tau A_t }

The endpoints are:

A_t^0=\epsilon, \qquad A_t^1=A_t

At $\tau=0$ , the sample is pure noise. At $\tau=1$ , it is the demonstrated action chunk. Intermediate values are partially noised action candidates. This straight line exists in generation space; it does not mean that the robot’s end effector follows a straight line in physical space.

5.3 Derive the paired training target

Rewrite the interpolation as:

A_t^\tau=\epsilon+\tau(A_t-\epsilon)

Differentiating with respect to $\tau$ gives:

\boxed{ \frac{dA_t^\tau}{d\tau}=A_t-\epsilon }

For each sampled pair $(\epsilon,A_t)$ , the target velocity is therefore:

u=A_t-\epsilon

In plain language, the target tells the model which direction and magnitude should be used to update the unfinished action candidate.

5.4 The action expert learns a conditional vector field

The model predicts:

v_\theta(A_t^\tau,o_t,\tau)

Its training loss is:

\boxed{ \mathcal L(\theta) = \mathbb E \left[ \left\| v_\theta(A_t^\tau,o_t,\tau) -(A_t-\epsilon) \right\|_2^2 \right] }

5.5 A necessary technical qualification

During training, each noise–demonstration pair supplies a target $A_t-\epsilon$ . At inference time, however, the model does not know a predetermined correct action $A_t$ .

Under the idealization of infinite data and mean-squared-error optimization, the learned field is the conditional vector field:

v^*(x,o,\tau) = \mathbb E \left[ A_t-\epsilon \mid A_t^\tau=x,\ o_t=o,\ \tau \right]

π₀ therefore learns distributional transport from Gaussian noise to an action distribution conditioned on the observation. It is not retrieving one particular training demonstration. Different initial noise samples can still yield different plausible actions; a conditional expectation vector field does not reduce the policy to one simple “average action.”

6. Inference: What the Ten Euler Steps Actually Do

6.1 Start from fresh noise

\hat A_t^{(0)}\sim\mathcal N(0,I)

At inference time, the current observation $o_t$ is available, but the correct action answer $A_t$ is not.

6.2 The learned ODE

\frac{d\hat A_t^\tau}{d\tau} = v_\theta(\hat A_t^\tau,o_t,\tau)

6.3 Discretize it with the forward Euler method

π₀ uses:

K=10, \qquad \delta=\frac{1}{K}=0.1

The update is:

\boxed{ \hat A_t^{(k+1)} = \hat A_t^{(k)} + \frac{1}{K} v_\theta \left( \hat A_t^{(k)},o_t,\frac{k}{K} \right) }

for $k=0,1,\ldots,9$ . The final output is:

\hat A_t=\hat A_t^{(10)}

6.4 Euler does not prove that ten steps must recover a real action

Consider a teaching example in which the velocity is assumed to be a known constant:

v_\theta=A_t-\epsilon

Then:

\hat A_t^{(k)} = \epsilon+\frac{k}{K}(A_t-\epsilon)

At $k=K$ :

\hat A_t^{(K)} = \epsilon+(A_t-\epsilon) =A_t

This algebra verifies only a known, constant-velocity path paired with one demonstration. It does not prove that ten steps are theoretically necessary or always sufficient in the real model, nor that every noise sample converges to a designated demonstration action. If the velocity were truly constant and the endpoint known, one step with $\delta=1$ would already reach it.

Multiple steps are useful because the learned vector field changes with the current action candidate, the observation, and $\tau$ . The system also has network-approximation and Euler-discretization error. Ten steps are therefore an engineering trade-off among generation quality, numerical accuracy, and inference cost—not a “ten-step convergence theorem.”

6.5 $H=50$ and $K=10$ mean different things

\boxed{ 10\text{ Flow updates} \longrightarrow 1\text{ action chunk containing 50 physical actions} }

$H=50$ is the number of physical time steps in the action chunk.
$K=10$ is the number of numerical integration steps used to generate that one chunk.

7. Training and Deployment, Side by Side

Training

real observation o_t and demonstrated action chunk A_t
→ sample noise ε
→ sample Flow time τ
→ construct intermediate action A_t^τ
→ train the action expert to predict A_t - ε
→ update model parameters

Deployment

current observation o_t and fresh noise
→ apply 10 Euler updates
→ obtain a 50-step action chunk
→ execute a prefix of the chunk
→ observe the real world again
→ replan

The demonstrated action is available as supervision during training. It is absent at deployment, where the policy must sample along the learned conditional vector field.

8. Beyond the Equations: Data and the Training Recipe

8.1 Data scale

The paper reports:

more than 10,000 hours of robot-manipulation data;
approximately 903M in-house timesteps;
seven robot configurations;
68 broad tasks;
open data representing roughly 9.1% of the training sampling mixture;
state and action interfaces padded to a maximum of 18 dimensions across robots.

The 903M figure counts timesteps, not complete trajectories. The 9.1% figure is a share of the sampling mixture, not a share of raw timesteps.

8.2 Broad pretraining

The pretraining data spans robots, tasks, objects, and scenes. It also includes imperfect actions, off-nominal states, and recoveries. Its purpose is not to make every behavior immediately expert-level, but to expand the range of states and behaviors that the policy has encountered.

8.3 High-quality post-training

Task-specific post-training uses more consistent, skilled, and targeted demonstrations to make behavior stable and fluid. A simple task may require about five hours of specialized data, while a complex task may require more than 100 hours.

π₀ does not eliminate task-specific data collection. It changes the role of that data from “learn the entire capability from scratch” to “calibrate and refine an existing capability.”

8.4 The complete recipe

\boxed{ \text{broad pretraining for capability coverage} \quad+\quad \text{high-quality post-training for behavioral proficiency} }

This is methodologically similar to broad pretraining followed by targeted adaptation. π₀’s task post-training, however, should not be equated directly with language-model RLHF or alignment.

9. What the Experiments Actually Establish

The paper evaluates the system at four levels:

Evidence level	Main result	Necessary qualification
Direct prompting	Across five tasks, π₀ obtained normalized progress scores of roughly 0.75–1.00 and substantially exceeded the paper’s baselines	These task families were present in pretraining; they were not strictly unseen tasks
Language following	The policy could use intermediate language instructions supplied by a person or a high-level VLM	π₀-small changes scale, initialization, and architecture together, so it is not a clean one-factor ablation
New-task adaptation	With 1, 5, or 10 hours of fine-tuning data, pretraining usually improved data efficiency	π₀ did not win on every task at every data point
Complex tasks	The paper demonstrated 5–20 minute tasks such as laundry handling, table clearing, and box assembly	“Above 50%” refers to partial progress scores, not full-task success rates

The 0.75–1.00 range is an approximate reading of Figure 7, not an exact numerical table reported by the paper.

9.1 The zero-shot wording in v1 versus v4

The initial v1 used the term zero-shot, while also stating that the five basic evaluation task families appeared in pretraining. The current v4 instead uses direct prompting and out-of-box.

The accurate interpretation is that the evaluated task instance received no task-specific post-training. It does not follow that the model had never encountered the related task family, robot, or behavior distribution.

9.2 Long tasks are not the same as fully autonomous planning

Some long-horizon tasks rely on intermediate instructions supplied by a person or a separate high-level VLM. π₀ primarily demonstrates a general low-level policy. A single π₀ model does not simultaneously provide long-horizon decomposition, persistent memory, success verification, safety judgment, and continuous low-level control.

10. What Was Genuinely Novel About π₀?

π₀ did not originate any one of the following ideas:

RT-2 introduced the VLA framing earlier;
Diffusion Policy explored continuous generative control earlier;
action chunking predates π₀;
Octo was an earlier cross-robot generalist policy;
OpenVLA earlier released a VLA built on an Internet-pretrained VLM.

π₀’s contribution is the system-level integration of:

a large pretrained VLM that preserves visual-language semantics;
a separate action expert for proprioceptive state and continuous actions;
Flow Matching that jointly generates high-frequency action chunks;
foundation-policy pretraining on more than 10,000 hours of cross-embodiment data;
high-quality post-training that turns broad capability into proficient task behavior;
extensive real-robot demonstrations involving deformable objects, bimanual coordination, and long tasks.

The paper itself frames the contribution as an integration and qualifies its novelty claim with “to our knowledge”: according to the authors, it was the first flow-matching VLA to generate high-frequency action chunks for dexterous control.

The most defensible historical assessment is:

π₀ is a defining starting point for Physical Intelligence’s π-series and the continuous action-expert VLA approach, but it is not the first work in VLA robotics or embodied intelligence as a whole.

11. My Interpretation: π₀ versus World-Model and WAM Approaches

This section is route-level analysis, not a conclusion established by the π₀ paper. WAM remains an emerging, nonstandard label; researchers do not yet share one definition of its boundaries or coupling mechanisms.

11.1 The key distinction is not “language medium versus video medium”

My initial intuition was that π₀ is more closely related to the LLM/VLM lineage, that VLA information is more language-like, and that WM/WAM information is more video- or latent-like. This identifies a difference in emphasis, but oversimplifies the intermediate representations.

A more precise distinction is:

π₀ inherits its semantic backbone from PaliGemma and therefore emphasizes objects, instructions, and task semantics;
the VLM does not pass readable language to the action expert, but a hidden context formed from images, language, and state;
WM/WAM approaches make future-world structure or predictive supervision materially participate in action learning.

A π₀-style direct policy learns:

p(A_t\mid o_t)

Its central question is: “Given the current observation, how should I act now?”

A world model may learn:

p(z_{t+1:t+H}\mid z_t,A_t)

Its central question is: “If these actions are taken, how might the world change?” Here $z$ may be an image, a visual latent, a state, or another world representation. It need not be a human-viewable video.

One teaching abstraction of a joint WAM is:

p(A_t,z_{t+1:t+H}\mid o_t)

This is not a canonical definition. Specific systems may use other factorizations, prediction only as training supervision, or an action-only output at deployment.

11.2 “Observe after acting” versus “simulate before acting” is a teaching contrast

π₀ follows a loop of:

observe → act directly → let the real world change → observe again

A world-model approach may instead use:

observe → predict possible futures internally → compare outcomes → choose an action

This is a useful representative contrast, not a universal rule. Not every world-model or WAM system explicitly renders video or searches multiple futures at deployment.

The absence of an explicit world model also does not mean that π₀ contains no physical knowledge. To generate effective actions from demonstrations, its weights may encode action-relevant regularities of contact, objects, and robot dynamics. The distinction is that π₀ does not train those regularities through a separately inspectable future-prediction objective, nor does it naturally expose a counterfactual simulator. It may know how to act without explicitly showing how the world will change after each action.

11.3 A shorter decision path may be faster, but speed is not guaranteed

A direct policy need not first generate a future world, evaluate candidate trajectories, and then select an action. Its decision path may therefore be shorter. However:

π₀ still performs ten action-expert Flow updates;
some WAMs use future prediction only during training and deploy an action-only policy;
engineering speed must be measured through end-to-end p50/p95 latency, control rate, hardware, and replanning behavior.

π₀ tends to trade a more direct action interface for efficient execution. WM/WAM approaches tend to trade richer dynamic representations for stronger consequence modeling. Their actual speed must be measured rather than inferred from the label.

11.4 The practical trade-off

Dimension	π₀-style direct VLA	WM/WAM-first approach
Primary learning target	conditional action distribution	coupling among world transitions, future representations, and actions
Main pre-action information	hidden context formed from the current observation	current observation plus future structure or predictive supervision
Must it generate video?	no	also no; it may use a latent or training-only supervision
Counterfactual rollouts	not explicitly provided	more naturally supports “what if we do this?”
Control interface	directly generates continuous action chunks	may generate jointly or use a policy/action-only head
Typical strength	clear semantic interface, short execution chain, continuous control	richer dynamics, planning, and consequence evaluation
Typical risk	may fail smoothly without explicit success verification	potentially higher compute, memory, data requirements, and model error

11.5 The approaches are more likely to merge than eliminate one another

My expectation is that world models are better suited to slow deliberation, long-horizon planning, consequence evaluation, and anomaly detection, while VLA action experts are better suited to fast, continuous low-level execution. This is an architectural inference from their complementary strengths, not a finding of the π₀ paper.

language and goal understanding
→ world-dynamics simulation
→ task and subgoal planning
→ continuous action generation
→ feedback from the real world

“VLA versus WAM” may therefore become less a product category than a distinction among capabilities inside one robotic system.

12. Greatest Strength and Most Durable Limitation

12.1 Greatest strength: a semantic-to-motor interface

π₀’s most durable contribution is neither a benchmark score nor the assumption that Flow Matching will remain the best generator forever. It connects the general semantic capabilities of a large VLM to a continuous action expert that can be replaced and extended.

Future systems may change the VLM backbone, action encoding, Flow solver, chunk length, or robot platform while preserving the division between a general semantic backbone and a continuous-control expert.

12.2 Greatest structural limitation: capability remains bounded by demonstration support

π₀ remains, at its core, an offline demonstration-driven conditional behavior-cloning system. It does not explicitly provide:

calibrated uncertainty about whether the state is outside its training distribution;
a refusal mechanism when the instruction or situation is not understood;
a verifier that determines whether an action actually completed the task;
persistent world state and long-term memory;
a mechanism for continual learning through online interaction.

It may therefore fail smoothly and confidently in unfamiliar states. More data can expand coverage, but it does not automatically create awareness of the unknown.

12.3 An enduring limitation of the paper’s evidence: weak causal attribution

The core 10,000-hour dataset is not fully available to third parties.
Baselines do not always share identical training budgets and action interfaces.
π₀ and π₀-small change parameter count, initialization, and architecture together.
Most conditions use roughly ten real-robot trials, and the paper does not report confidence intervals.
Complex tasks use author-designed partial-progress metrics.

openpi now provides public code and base weights, so describing π₀ as “entirely closed source” is no longer accurate. The original 10,000-hour pretraining run, however, still cannot be reproduced in full by a third party.

The paper provides strong evidence that the complete recipe worked in the authors’ environments. It does not cleanly isolate how much of the gain came from the VLM, Flow Matching, data scale, post-training, or systems engineering.

13. Common Misconceptions

π₀ is formally short for “Physical Intelligence Zero.” No. It should simply be called π₀ or pi-zero.
π₀ is a world model. No. It does not explicitly predict future images, states, or rewards.
The VLM first emits a sentence that is passed to the action model. No. The action expert reads internal hidden context.
$H=50$ means that the action is 50-dimensional. No. It denotes 50 physical time steps.
Ten Flow updates generate ten action chunks. No. The ten updates jointly generate one action chunk.
Forward Euler proves that ten steps must reach the correct action. No. Euler is a numerical solver, and ten steps are an engineering choice.
50 Hz means that the full visual model replans 50 times per second. No. It is the action-command rate; the system replans roughly every 0.5–0.8 seconds.
Direct prompting means that the model has never seen the task. No. The basic task families appeared in pretraining.
π₀ alone performs all planning for the long tasks. Not in every case. Some tasks use intermediate instructions from a person or a high-level VLM.
The experiments prove that Flow Matching alone caused π₀’s success. No. The paper primarily validates the complete system recipe.

14. One-Minute Recap

π₀ is Physical Intelligence’s first-generation generalist robot policy. It uses a roughly 3B-parameter PaliGemma backbone to encode images and language, then a roughly 300M-parameter action expert to generate an action chunk containing 50 continuous physical actions through Conditional Flow Matching. During training, the model constructs a straight path between Gaussian noise and a demonstrated action and learns how an intermediate action candidate should change. During inference, it begins from fresh noise, applies ten forward Euler updates, executes only part of the resulting chunk, and then observes the real world again. π₀’s central innovation is not one isolated equation. It is the integration of VLM semantics, a continuous action expert, more than 10,000 hours of cross-embodiment pretraining, and high-quality post-training into a foundation-policy recipe. The fundamental difference from a world model is that π₀ directly learns a conditional action distribution, whereas a world model explicitly learns how the world may change. Over time, the two approaches are more likely to combine across planning and execution layers.

15. Review Questions

If you can answer these ten questions, you have captured the core of π₀:

What are π₀’s inputs and outputs?
What do $a_t$ , $A_t$ , $H$ , $d$ , and $D$ represent?
How does an action chunk differ from a discrete action token?
Why does generation begin from Gaussian noise?
Why is the paired training target $A_t-\epsilon$ ?
Why is no known correct $A_t$ available at inference time?
What does forward Euler do in this system, and what does it not establish?
Why are $H=50$ and $K=10$ conceptually independent?
Why is π₀’s innovation a system recipe rather than one Flow equation?
How do the learning objectives of π₀ and world-model/WAM approaches differ?

Primary Sources

Core paper: π₀ v4, π₀ v1
Author materials: PI π₀ project page, openpi
Methodological lineage: PaliGemma technical report, Flow Matching, Rectified Flow
Prior work: RT-2, Diffusion Policy, Octo, OpenVLA
Subsequent boundaries: π₀.5, Real-Time Chunking
WAM route references: WAM Survey, VPP, DreamZero, Fast-WAM, GigaWorld-Policy
Paper index: Hugging Face paper page

Versioning and Verification Notes

Equations, architecture, data, and experimental claims primarily follow π₀ arXiv v4.
Historical zero-shot wording follows v1, with the v4 revision stated explicitly.
The paper’s derivation uses $\tau=0$ for noise and $\tau=1$ for action. The current openpi code uses the opposite time convention— $t=1$ for noise, $t=0$ for action, and $dt<0$ . The substitution $t=1-\tau$ makes them equivalent; this is not a contradiction.
The main text therefore keeps the action dimension generic as $d$ . Eighteen dimensions are the paper’s padded cross-robot data interface; the current open-source Pi0Config defaults to action_dim=32. They describe different implementation layers and must not be conflated.
The openpi repository was confirmed as public through the GitHub connector on August 2, 2026.
The Hugging Face page uses a snapshot verified on August 1, 2026 and serves only as metadata and an ecosystem index.
“My interpretation” is explanatory analysis and should be read separately from claims established by the paper.

π₀：机器人怎样把“看懂任务”变成连续动作π₀: How a Robot Turns Understanding into Continuous Action