12 个 case，同一份输入

所有时间均相对 episode 起点，25Hz tick；不是原始打包视频的绝对时间。每次输入是 frame0 锚点 + 8 个历史/当前帧。每个 case 的 JSON 同时提供相机原始路径、视频偏移、8 段时间表、精确 prompt、seed、归一化状态和有效维度。

完整 JSON · CSV · 复现说明 · Python 输入加载器

Case	数据集	原始 episode ID	处理后 episode / shard	预测起点	历史范围	下载
case15	robocoin_lerobot_v3/robocoin_agilex_decoupled_hd	robocoin_lerobot_v3:Agilex_Cobot_Magic_scoop_coffee_beans/episode_000128	515 / c000-f001	16.28s	9.28s–16.28s + frame0	JSON · 初始输入
case13	robocoin_lerobot_v3/robocoin_agilex_cobot_magic_with_front_chest	robocoin_lerobot_v3:Agilex_Cobot_Magic_connect_block/episode_000003	3 / c000-f000	10.28s	3.28s–10.28s + frame0	JSON · 初始输入
case21	robocoin_lerobot_v3/robocoin_galaxea_r1_lite_stereo	robocoin_lerobot_v3:Galaxea_R1_Lite_mix_red_yellow_left_large_test_tube/episode_000034	690 / c000-f000	25.56s	18.56s–25.56s + frame0	JSON · 初始输入
case14	robocoin_lerobot_v3/robocoin_agilex_decoupled	robocoin_lerobot_v3:Cobot_Magic_pour_drink/episode_000009	273 / c000-f000	18.72s	11.72s–18.72s + frame0	JSON · 初始输入
case19	robocoin_lerobot_v3/robocoin_ai2_alphabot2_s28	robocoin_lerobot_v3:alpha_bot_2_press_the_button_a/episode_000004	41 / c000-f000	4.80s	0.00s–4.80s + frame0	JSON · 初始输入
case23	robocoin_lerobot_v3/robocoin_realman_aidal	robocoin_lerobot_v3:RMC-AIDA-L_food_packaging/episode_000033	434 / c000-f002	36.72s	29.72s–36.72s + frame0	JSON · 初始输入
case03	agibotworld_2026_lerobot_v3/s193a44	agibotworld_2026_lerobot_v3:ImitationLearning__CommercialSpaces__task_4542__465501_467318#25	25 / c000-f000	15.76s	8.76s–15.76s + frame0	JSON · 初始输入
case17	robocoin_lerobot_v3/robocoin_agilex_decoupled_with_high_realsense	robocoin_lerobot_v3:Cobot_Magic_make_hamburger/episode_001514	1514 / c000-f001	2.24s	0.00s–2.24s + frame0	JSON · 初始输入
case05	agibotworld_2026_lerobot_v3/s216a44_rl	agibotworld_2026_lerobot_v3:ReinforcementLearning__Industry__task_7093__PolicyRollout__20758141_20758461#18	981 / c000-f021	30.40s	23.40s–30.40s + frame0	JSON · 初始输入
case04	agibotworld_2026_lerobot_v3/s216a44	agibotworld_2026_lerobot_v3:ReinforcementLearning__Home__task_12173__ExpertDemonstration__12105429_14973988#31	31 / c000-f000	25.24s	18.24s–25.24s + frame0	JSON · 初始输入
case01	agibotworld_2026_lerobot_v3/s169a40	agibotworld_2026_lerobot_v3:ImitationLearning__CommercialSpaces__task_3400__313498_314085#11	11 / c000-f000	17.52s	10.52s–17.52s + frame0	JSON · 初始输入
case02	agibotworld_2026_lerobot_v3/s169a40_hires	agibotworld_2026_lerobot_v3:ImitationLearning__CommercialSpaces__task_3477__250434_252697#10	10 / c000-f000	13.76s	6.76s–13.76s + frame0	JSON · 初始输入

图像包是原始推理使用的无损 RGB 复合画面，视频播放器展示的 MP4 仅供观看。GT 反馈每段读取真实输入；预测反馈后续必须使用待测模型自己的输出，不能偷用后续 GT。