论文记录 · 证据边界公开

π₀: A Vision-Language-Action Flow Model for General Robot Control

Robotics: Science and Systems XXI · 2025-06-21 · 通用或轻量 VLA

论文元数据

作者: Kevin Black、Noah Brown、Danny Driess、A. Esmail、Michael Equi、Chelsea Finn、Niccolo Fusai、Lachy Groom、Karol Hausman、Brian Ichter、Szymon Jakubczak、Tim Jones、Liyiming Ke、Sergey Levine、Adrian Li-Bell、Mohith Mothukuri
机构: 机构尚未补齐
发布日期精度: 日
研究方向: 通用或轻量 VLA
机器人: 未关联具体机器人
DOI: 10.15607/rss.2025.xxi.010
arXiv: 2410.24164
引用元数据: 140
记录状态: included
质量检查: 无自动质量红旗

工程资源状态

代码: 已核验开放
权重: 已核验可访问（使用限制见许可证）
数据: 已核验可访问（使用限制见许可证）
许可证: 许可证条款已核验（不等于可商用）

已终审工程资源

官方代码 · 许可证 Apache-2.0 · 最近核验 2026-07-19
依赖：Ubuntu 22.04、Python 3.11 or newer、JAX CUDA 12 build、Flax、Orbax Checkpoint、LeRobot pinned dependency、optional PyTorch and websocket remote-inference paths；硬件/传感器：more than 8 GB VRAM for inference、more than 22.5 GB VRAM for LoRA fine-tuning、more than 70 GB VRAM for full fine-tuning、RGB cameras, proprioception and language prompt；实机证据：Official π0 paper evaluates more than 20 physical fine-tuning tasks with ten-trial scoring, and the repository includes ALOHA and DROID integration guides. No independent runtime reproduction was performed in this review.；开放范围/限制：Apache-2.0 covers repository code. The current main branch also contains π0-FAST and π0.5; results and artifacts must remain versioned rather than attributed wholesale to the original π0 paper.
官方数据 · 许可证 Apache-2.0 for this public example dataset only · 最近核验 2026-07-19
依赖：LeRobot-compatible dataset tooling、openpi data conversion and training configuration、dataset-specific download and preprocessing；硬件/传感器：ALOHA-class bimanual setup for corresponding physical collection、local storage and GPU compute for fine-tuning；实机证据：The official repository uses selected public datasets for adaptation examples, but Physical Intelligence's complete 10,000-hour seven-configuration π0 training mix and all paper post-training sets are not public.；开放范围/限制：Apache-2.0 applies to this named public dataset only. This record must not be displayed as the complete π0 paper training corpus, which remains substantially unreleased.
官方权重 · 许可证 Gemma Terms of Use apply; no standalone pi0_base checkpoint license was found · 最近核验 2026-07-19
依赖：openpi repository、JAX and Flax、Orbax Checkpoint、PaliGemma/Gemma-derived model components；硬件/传感器：more than 8 GB VRAM for documented inference、paper reports 73 ms on-board or 86 ms off-board for three-camera input on RTX 4090、robot-specific action-space conversion and policy server integration；实机证据：The π0 family is evaluated on physical UR5e, Franka, Trossen, ARX and mobile configurations in the official paper; this review verified anonymous bucket listing but did not download or execute the 11.189 GiB base checkpoint.；开放范围/限制：Do not inherit Apache-2.0 from the code repository onto the checkpoint. The repository includes Gemma terms, and no standalone license object was found in the public checkpoint bucket root.

“官方”表示归属与论文关系已核验；仍应按所列依赖、硬件和许可证复现。

原始来源

元数据审计 · 1 条

LOW · 仓库与论文标题关系需人工复核
人工核对仓库 README、论文引用和作者归属，排除同名误配。

审计问题是修复待办，不代表论文结论无效；未经原始来源终审不会自动改库。

最近核验：2026-07-19。链接存在不等于官方仓库、可运行、已开放权重/数据或许可证允许使用。