论文记录 · 证据边界公开

Generalizable Humanoid Manipulation with 3D Diffusion Policies

IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2025; author-stated on arXiv) · 2024-10-14 · 人形全身运动跟踪

论文元数据

作者: Yanjie Ze、Zixuan Chen、Wenhao Wang、Tianyi Chen、Xialin He、Ying Yuan、Xue Bin Peng、Jiajun Wu
机构: 机构尚未补齐
发布日期精度: 日
研究方向: 人形全身运动跟踪
机器人: 未关联具体机器人
DOI: 10.48550/arXiv.2410.10803
arXiv: 2410.10803
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: 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 MIT · 最近核验 2026-07-19
依赖：Python 3.8、PyTorch 2.1.0、Open3D、torch-geometric、diffusers 0.11.1、zarr 2.12.0、pyrealsense2 2.54.2.5684、companion Humanoid-Teleoperation stack；硬件/传感器：RTX 4090 24 GB for documented training、Fourier GR-1 with Inspire Hands、Intel RealSense L515、Apple Vision Pro for teleoperation；实机证据：Official paper reports more than 2,000 physical policy rollouts on Fourier GR-1 across Pick and Place, Pour and Wipe; no independent runtime reproduction was performed in this review.；开放范围/限制：The documented deployment is hardware-specific and requires companion communication code. The lower body is disabled and the robot uses a cart, so this is not bipedal loco-manipulation evidence.
官方数据 · 许可证 unknown · 最近核验 2026-07-19
依赖：Improved-3D-Diffusion-Policy dataset loader、zarr 2.12.0、Open3D、pyrealsense2 and L515 calibration for equivalent new data；硬件/传感器：RealSense L515 RGB-D/LiDAR camera for corresponding collection format、Fourier GR-1 and Inspire Hands for the paper's full task data、Apple Vision Pro teleoperation stack for collecting equivalent demonstrations；实机证据：The official project reports 300 training episodes across three tasks and releases full-task archives in the companion Drive folder, but this review did not unpack them to independently verify counts, synchronization or calibration.；开放范围/限制：This distinct URL is the official training-data example used by the README. Full-task archives share the checkpoint folder, but neither the example nor those archives expose a standalone data license; do not infer MIT from the code repository.
官方权重 · 许可证 unknown · 最近核验 2026-07-19
依赖：Improved-3D-Diffusion-Policy repository、PyTorch 2.1.0、Open3D and torch-geometric、task-specific configuration and robot API integration；硬件/传感器：Fourier GR-1、two Inspire dexterous hands、RealSense L515 RGB-D/LiDAR camera、CPU-only onboard inference reported at 15 Hz；实机证据：Three official task checkpoint archives are anonymously listed and correspond to the paper's physical GR-1 evaluation, but the archives were not fully downloaded or executed in this review.；开放范围/限制：The Drive folder exposes three checkpoint ZIP files, but no standalone model license, version manifest or published checksums were visible. Public downloadability does not imply MIT or commercial redistribution rights.

“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

最近核验：2026-07-20。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。