论文记录 · 证据边界公开

Jetson-PI: Towards Onboard Real-Time Robot Control via Foresight-Aligned Asynchronous Inference

arXiv · 2026-07-14 · 低算力 VLA 实时推理与边缘运行时

论文元数据

作者: Zebin Yang、Qi Wang、Yunhe Wang、Xiurui Guo、Bo Yu、Shaoshan Liu、Jiafeng Xu、Hao Dong、Meng Li
机构: Peking University、AIRS、PrimeBot Research Institute
发布日期精度: 日
研究方向: 低算力 VLA 实时推理与边缘运行时
机器人: 未关联具体机器人
DOI: 未披露
arXiv: 2607.12659
引用元数据: 0
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: 未核验
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 for project code; bundled Gemma, openpi and LIBERO components retain separate terms · 最近核验 2026-07-21
依赖：Ubuntu 22.04、Python 3.11、JAX 0.5.3、CUDA 12.x、OpenPI / π0.5、LIBERO、MuJoCo/EGL or Xvfb、Orbax；硬件/传感器：At least 48 GB GPU memory recommended for batch-16 three-stage training、NVIDIA A100 used in the paper、NVIDIA Jetson Orin/Thor or desktop NVIDIA GPU for deployment/evaluation；实机证据：The paper reports X2-W cloth folding at 15 Hz with three cameras and 10 trials per subtask; the repository itself exposes only the demo media and LIBERO release, not the private X2-W robot stack.；开放范围/限制：Verified at commit 35de6060003f6260df293771ef083ff1033c99b6. The repository contains 160 tracked files, including 101 Python files; Python compileall passed. It open-sources LIBERO training/evaluation and policy serving, but not the X2-W physical dataset/controller. There is no formal GitHub release, and upstream component licenses remain separate.
官方代码 · 许可证 MIT for the repository, with bundled llama.cpp and third-party notices retained · 最近核验 2026-07-21
依赖：CMake、C/C++ toolchain、CUDA for Jetson/NVIDIA acceleration、Python/NumPy for managed client、PI0 or PI0.5 source checkpoint、separate LLM/action and SigLIP GGUF conversion；硬件/传感器：NVIDIA Jetson Orin in the reported MAXN latency benchmark、Jetson Thor or CUDA desktop GPU、CPU builds are supported but not the reported real-time target；实机证据：The runtime accompanies the X2-W/Jetson Orin cloth-folding experiment; no independent robot deployment at the fixed commit was found.；开放范围/限制：Verified at commit 04de0b3bdee3f45aa848cc92b51fffe0655317bb with 3,031 tracked files. The repository supports PI0/PI0.5, persistent HTTP, Python and FlashRT interfaces. No preconverted GGUF weights or formal release are supplied, so users must patch source model metadata, split the vision tower and convert two matching files.
官方权重 · 许可证 Mixed terms: ModelScope declares Gemma; pi05_libero retains openpi and Gemma terms; the future correction module follows the Jetson-PI Apache-2.0 project terms · 最近核验 2026-07-21
依赖：Jetson-PI repository、JAX/Orbax loader、OpenPI π0.5 configuration、LIBERO norm statistics、approximately 12.60 GB local storage；硬件/传感器：29 downloadable blobs totaling 12,603,238,852 bytes、GPU with sufficient memory for π0.5 inference、GGUF conversion required for Jetson-PI-Edge；实机证据：The published bundle is a LIBERO checkpoint plus the future correction module; it is not the paper's X2-W XR-1 physical checkpoint and does not reproduce the physical trial without a robot-specific policy and controller.；开放范围/限制：Verified at latest ModelScope revision 95207c5790aacbb3722b009111bbca1578993b3b. The 29 blobs include the π0.5-LIBERO backbone and a step-65000 future correction module. The model page is publicly accessible, but its card contains a stale Hugging Face download placeholder; use the ModelScope snapshot command documented in the official code repository. License obligations must be applied per component.

“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。

每日候选人工终审

Jetson-PI 同时处理异步 VLA 的感知—执行错位和边缘端系统瓶颈：未来表征校正、置信度调度、CUDA 图复用、GPU 常驻缓存与 flow unroll 使 Orin 上控制频率从 0.70 Hz 提升到 6.06 Hz，并在 LIBERO 长延迟条件下保持 92.5% 平均成功率，属于低算力 VLA 部署 2.0、可能成为新主流。X2-W 折衣三子任务各 10 次达到 10/10、8/10、9/10；训练/评测代码、llama.cpp 边缘运行时和 12.60 GB 权重均已核验开放。局限是训练仍需 48 GB 显存/A100，边缘运行时要自行转换 GGUF，权重含 Gemma/openpi 混合条款，且独立复现尚未出现。

实机证据：quantified_x2_w_cloth_folding_on_jetson_orin；折衣分拣放三阶段各 10 次：RTX 4090 为 10/10、7/10、10/10；Orin naive async 为 6/10、0/10、5/10；Jetson-PI 为 10/10、8/10、9/10，动作执行 15 Hz。LIBERO 与系统表另分别报告 50 次/任务和 Orin 0.70→6.06 Hz。

硬件/本体：PrimeBot X2-W with head and two wrist cameras at 224×224、NVIDIA Jetson Orin、RTX 4090 comparison deployment

红旗与边界：完整三阶段训练推荐至少 48 GB 显存，论文训练用 A100，低算力优势主要发生在部署端；ModelScope 权重不是单一 Apache-2.0：π0.5/Gemma 与校正模块条款必须分开；Edge 仓库没有预转换 GGUF，用户需补配置、拆分并转换两个模型文件；实机只验证一个折衣流程、每子任务 10 次，跨机器人泛化仍未量化；两个代码仓均无正式 release，主仓仍有权重镜像和封闭评测相关开放事项；尚未发现独立团队固定版本复现

查看资源 · code · 状态 verified_open_fixed_commit · 许可证 Apache-2.0 for project code; Gemma, openpi and LIBERO terms remain separate
固定提交 35de6060003f6260df293771ef083ff1033c99b6，160 个受控文件、101 个 Python 文件；Python 静态编译通过。包含三阶段训练、LIBERO 评测、策略服务和 50 次/任务默认配置。
查看资源 · code · 状态 verified_open_fixed_commit · 许可证 MIT with bundled third-party licenses
固定提交 04de0b3bdee3f45aa848cc92b51fffe0655317bb，3,031 个受控文件；支持 PI0/PI0.5、CUDA/CPU、HTTP、Python 与 FlashRT 接口。需将策略和 SigLIP 分别转换为 GGUF。
查看资源 · weights · 状态 verified_open_modelscope_revision · 许可证 Mixed: ModelScope metadata marks Gemma; pi05_libero retains openpi/Gemma terms; future correction module follows Jetson-PI Apache-2.0
固定最新 revision 95207c5790aacbb3722b009111bbca1578993b3b，29 个 blob、12,603,238,852 bytes；包含约 12 GB π0.5-LIBERO 和约 157 MB future correction module。
data · 状态 uses_upstream_libero_no_new_dataset_release · 许可证 LIBERO and upstream terms
项目使用上游 LIBERO 数据，没有发布新的物理折衣数据或专属训练集。

论文入库不等于代码、权重或数据开放；“承诺开放”“声称开放”和“已核验可下载”严格分开。

原始来源

最近核验：2026-07-21。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。