iUMI_Inference / docs /ROLLOUT_VALIDATION.md
Crobotia's picture
Add OpenPI and FLUX demo videos, sample rollouts and expanded model card
bcc05dc verified
|
Raw History Blame Contribute Delete
3.38 kB

Demo 与 rollout 发布检查 / Demo and rollout release checks

发布日期 / Release date: 2026-10-27. Machine-readable results: rollout_validation.json.

本次在上传前,对新增材料重新执行以下检查:

检查 结果
两套数据原有 SHA-256 清单 全部通过,各 13 个文件
帧级 Parquet / episode 索引 OpenPI 6,540 行、FLUX 30,619 行,各 1 个 episode;索引连续、时间递增
状态与动作 均为有限 16 维数值;除末帧外满足下一帧状态关系
四路 rollout 视频 AV1、960×960、30 Hz,无音轨;视频元数据及包计数均与对应帧数一致
四个 IMU CSV 所有数值有限、时间戳递增、行数与导出记录一致
两个机器人 H5 可打开;记录数据集形状和类型
两段 demo H.264 视频流、无音轨;OpenPI 255.4 秒,FLUX 269.7 秒

完整 rollout 视频的本次检查使用容器元数据和包计数,没有重新解码全部视频帧。H5 可读性检查不等于运动学或控制正确性验证。源转换器的重复选帧计数予以保留:OpenPI 左/右 7/6,FLUX 左/右 30/31;跳帧计数均为 0。这些是源记录,不是本次重新运行的重复帧检测。

此次上传保留原始数据和视频字节,排除 .DS_Store。每套数据的局部清单与根目录合并清单分别从对应目录执行 shasum -a 256 -c checksums.sha256。根目录清单包含现有源码/权重记录与新增材料;未重新下载或校验原有模型大文件。

演示是展示剪辑,不能与完整 rollout 按帧号对齐。样例未附确切 checkpoint 哈希和录制时采样参数。本次未重新运行推理、实机控制、完整运动质检或任务成功率评测,也未将人工演示数据的 QC 结论应用到 rollout。

Before upload, the added assets were checked for SHA-256 integrity (13 existing entries per dataset), Parquet row/index consistency, increasing timestamps, finite 16-D states/actions and next-state labels. Four rollout video streams have matching container frame/packet counts at 960×960 / 30 Hz AV1 with no audio. All four IMU CSVs contain finite values and increasing times, with row counts matching export metadata. Both H5 files open successfully; dataset shapes and dtypes are recorded. The two H.264 demos contain no audio and last 255.4 / 269.7 seconds.

Rollout videos were inspected via container metadata and packet counts, not a new full-frame decode. H5 readability does not validate kinematics or control. Source repeated-frame-selection counts are retained (OpenPI left/right 7/6; FLUX 30/31; no skipped frames); duplicate-frame detection was not rerun.

Original data/video bytes and dataset-local manifests are preserved. .DS_Store files are excluded. Run shasum -a 256 -c checksums.sha256 from the appropriate dataset or repository root. The combined manifest retains existing source/weight entries; existing large model files were not downloaded or revalidated in this update.

Presentation edits cannot be frame-joined to the complete rollouts. Exact recording checkpoint hashes and sampling settings were not supplied. This upload does not rerun model inference, robot control, complete motion QC or task-success evaluation, and does not apply human-demonstration QC findings to these samples.