video video 9 12.4 | label class label 2
classes |
|---|---|
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
0video.cam_high | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist | |
1video.cam_right_wrist |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
J_VLA_v0 Task-Complete 数据集说明
概览
本目录是 /public/home/zhuyuchen530/J_VLA_v0 的任务完成版副本。原数据在
蓝色方块已经放入绿色碗后,还包含机械臂撤回并接近初始位姿的较长动作。本次
转换将 episode 终点改为“任务成功已经明确可见”,不再等待机械臂完全回位。
- 源目录:
/public/home/zhuyuchen530/J_VLA_v0 - 输出目录:
/public/home/zhuyuchen530/J_VLA_v0_task_complete - 转换脚本:
/public/home/zhuyuchen530/projects/le-wm/scripts/trim_j_vla_to_task_success.py - 数据格式:LeRobot v2.1
- 任务:
Put the blue cube into the green bowl - 帧率:30 FPS
- Episodes:20
- 相机:
video.cam_high、video.cam_right_wrist - 帧数:9,615 -> 6,276
- 删除:3,339 帧,约 34.7%
源数据目录不会被脚本写入或覆盖。
任务完成检测
当前数据是单一 pick-and-place 任务,因此使用抓手释放作为可复现的任务完成 候选事件:
- 读取
observation.state[..., 6]的连续抓手状态。 - 从每条 episode 的 35% 之后开始搜索,排除起始阶段张开抓手的动作。
- 将
gripper[t+1] - gripper[t] > 0.00015视为张开变化。 - 将间隔不超过 15 帧的张开变化合并为同一个事件。
- 选择第一个累计张开量不小于
0.002的事件。 - 在释放结束后额外保留 20 帧,约 0.67 秒。
额外保留的 20 帧用于记录方块已经稳定落入碗中,以及机械臂开始撤离。随后 回到初始位姿的动作被删除。20/20 条 episode 的主相机裁剪终点均已人工审查。
该规则依赖“释放物体即完成任务”的语义,不能不经检查地用于推、拉、开门或 无需松开抓手的任务。
每条 Episode 的长度
| Episode | 原长度 | 释放区间 | 新长度 | 删除 |
|---|---|---|---|---|
| 0 | 500 | 262-288 | 309 | 191 |
| 1 | 464 | 285-295 | 316 | 148 |
| 2 | 500 | 264-273 | 294 | 206 |
| 3 | 458 | 283-295 | 316 | 142 |
| 4 | 420 | 250-275 | 296 | 124 |
| 5 | 458 | 239-249 | 270 | 188 |
| 6 | 405 | 260-269 | 290 | 115 |
| 7 | 500 | 258-266 | 287 | 213 |
| 8 | 500 | 302-310 | 331 | 169 |
| 9 | 500 | 282-288 | 309 | 191 |
| 10 | 466 | 253-264 | 285 | 181 |
| 11 | 472 | 299-308 | 329 | 143 |
| 12 | 500 | 303-320 | 341 | 159 |
| 13 | 500 | 341-350 | 371 | 129 |
| 14 | 500 | 262-272 | 293 | 207 |
| 15 | 500 | 283-298 | 319 | 181 |
| 16 | 500 | 286-295 | 316 | 184 |
| 17 | 484 | 261-284 | 305 | 179 |
| 18 | 500 | 307-318 | 339 | 161 |
| 19 | 488 | 330-339 | 360 | 128 |
更完整的参数、候选释放事件和生成记录见
conversion_manifest.json;简要转换报告见 CONVERSION_REPORT.md。
转换内容
脚本对所有时序数据使用同一个裁剪长度:
- 截断两个相机的 MP4,并重新编码为 H.264、
yuv420p。 - 截断对应 parquet 中的
observation.state、action和索引字段。 - 重置 episode 内的
frame_index与timestamp。 - 重建跨 episode 连续的全局
index。 - 更新
meta/episodes.jsonl和meta/info.json。 - 重建
meta/steps_data_index.pkl。 - 使用裁剪后的全部帧重新计算
meta/stats.json。 - 同步重建
meta/stats_gr00t.json。
视频先在节点本地 /tmp 完成封装,再顺序复制到输出目录,以避免 /public
文件系统上的 MP4 随机写问题。整个数据集先写入同级临时目录,全部成功后再
原子 rename。目标目录已存在时脚本会直接退出,防止误覆盖。
复现转换
因为当前目录已经存在,复现时应指定新的输出目录:
cd /public/home/zhuyuchen530/projects/le-wm
source .venv/bin/activate
python scripts/trim_j_vla_to_task_success.py \
--source /public/home/zhuyuchen530/J_VLA_v0 \
--output /public/home/zhuyuchen530/J_VLA_v0_task_complete_rebuild \
--visual-audit-status "manual review required"
关键参数可通过以下选项调整:
--release-search-fraction
--positive-delta
--merge-gap
--min-release-gain
--post-success-frames
LeWM 训练
对应的数据配置为:
/public/home/zhuyuchen530/projects/le-wm/config/train/data/j_vla_v0_task_complete_fs2.yaml
它保持原 J_VLA 设置:frameskip=2、7D 连续 action/state、min_max
归一化,以及 episode-level 90/10 train/val 划分。使用已有两阶段微调脚本:
cd /public/home/zhuyuchen530/projects/le-wm
DATA_ROOT=/public/home/zhuyuchen530/J_VLA_v0_task_complete \
RUN_NAME=j_vla_task_complete_run01 \
scripts/run_j_vla_v0_finetune.sh
方法依据
- DROID 使用任务 success 标注区分可用于训练的成功示范: https://www.jiajunwu.com/papers/droid_rss.pdf
- LIV 使用语言条件视觉表示追踪任务视频中的进度: https://proceedings.mlr.press/v202/ma23b/ma23b.pdf
- RoboCLIP 和 RoboReward 展示了使用视频模型或 VLM 构造任务进度奖励的 路径:https://arxiv.org/abs/2310.07899、 https://arxiv.org/abs/2601.00675
对于当前仅 20 条、单任务且释放事件明确的数据,机器人状态检测加全量人工 审查比引入 learned reward 更直接,也更容易复现。
- Downloads last month
- 54