[Bug] 4 corrupt motion files in `data/motions/rpo_lab` (unreadable pickle)（rpo_lab 中有 4 个损坏的动作文件）

[mecha-d]

Affected files (paths relative to robolab):

• data/motions/rpo_lab/run_start_180_R_001__A345.pkl (112,194 bytes)
• data/motions/rpo_lab/run_start_180_R_001__A345_M.pkl (112,194 bytes)
• data/motions/rpo_lab/run_stop_180_R_001__A345.pkl (149,746 bytes)
• data/motions/rpo_lab/run_stop_180_R_001__A345_M.pkl (149,746 bytes)

Environment: verified at commit efb32c6; the corruption is in the
committed git blobs themselves (see below), so it reproduces on any
OS/Python.

Expected: pickle.load() returns the motion dict (fps, root_pos,
root_rot, dof_pos, key_body_pos) as it does for the other 26 clips.

Actual: pickle.load() fails on all four with:

_pickle.UnpicklingError: invalid load key, '\x05'.

The files begin with a valid pickle protocol-4 header (\x80\x04\x95) but the
stream is corrupt partway through. Note each _M file is byte-identical in
size to its non-_M partner, which suggests a single failed export/copy step
produced both.

Reproduction:

import pickle
with open("data/motions/rpo_lab/run_start_180_R_001__A345.pkl", "rb") as f:
    pickle.load(f)  # UnpicklingError: invalid load key, '\x05'.

Verified not a download problem: the on-disk bytes match the committed git
blobs exactly (git status clean, git cat-file blob header identical to
disk), and no LFS is involved — the corruption is in the repository content
itself. All other 26 clips in rpo_lab load and parse fine.

Impact: any pipeline that iterates over rpo_lab/*.pkl (e.g. AMP dataset
loading) crashes unless these four files are skipped.

Suggested fix: re-export these clips from the source data and recommit, or
remove them from the dataset.

---

中文版本（点击展开）/ Chinese version

受影响的文件（路径相对于 robolab）：

• data/motions/rpo_lab/run_start_180_R_001__A345.pkl（112,194 字节）
• data/motions/rpo_lab/run_start_180_R_001__A345_M.pkl（112,194 字节）
• data/motions/rpo_lab/run_stop_180_R_001__A345.pkl（149,746 字节）
• data/motions/rpo_lab/run_stop_180_R_001__A345_M.pkl（149,746 字节）

环境： 在提交 efb32c6 上验证；损坏存在于 git 提交的 blob 本身（见下文），
因此在任何操作系统/Python 版本上都可复现。

预期行为： pickle.load() 返回动作字典（fps、root_pos、root_rot、dof_pos、
key_body_pos），与其余 26 个文件一致。

实际行为： 这四个文件用 pickle.load() 读取时全部失败，报错：

_pickle.UnpicklingError: invalid load key, '\x05'.

文件开头是有效的 pickle protocol-4 头（\x80\x04\x95），但数据流在中途已损坏。
另外，每个 _M 文件与对应的非 _M 文件字节大小完全相同，推测是同一次失败的
导出/复制操作同时产生了这两组文件。

复现方法：

import pickle
with open("data/motions/rpo_lab/run_start_180_R_001__A345.pkl", "rb") as f:
    pickle.load(f)  # UnpicklingError: invalid load key, '\x05'.

已确认不是下载问题： 磁盘上的字节与 git 提交的 blob 完全一致（git status
干净，git cat-file blob 的文件头与磁盘相同），也未使用 LFS——损坏存在于仓库
内容本身。rpo_lab 中其余 26 个动作文件均可正常读取和解析。

影响： 任何遍历 rpo_lab/*.pkl 的流程（例如 AMP 数据集加载）都会崩溃，
除非跳过这四个文件。

建议修复： 从源数据重新导出这些动作并重新提交，或将它们从数据集中移除。