File size: 2,763 Bytes
06f9561
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
{
  "format": "hfd_multiturn_auto_infer_scenario_v1",
  "name": "interrupt_grab_wave_walk_left_zh",
  "interaction_mode": "interruption",
  "data_split": "train",
  "split_seed": 42,
  "val_ratio": 0.1,
  "split_source_namespace": "seed",
  "description": "三轮训练集指令;第二、三条语音分别在上一动作开始约 4.0/4.8 秒后到来,用于检查因果打断、连续 KV 和连续 RTC。",
  "required_semantics_version": 20,
  "initial_idle_steps": 5,
  "initial_idle_seconds": 2.0,
  "motion_frames_per_turn": 450,
  "motion_seconds_per_turn": 9.0,
  "inter_turn_background_steps_by_turn": [
    10,
    12
  ],
  "inter_turn_background_seconds_by_turn": [
    4.0,
    4.8
  ],
  "terminal_background_steps": 5,
  "terminal_background_seconds": 2.0,
  "turns": [
    {
      "turn_index": 1,
      "source_motion_filename": "grab_walk_ff_180_001__A548",
      "audio_path": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/buwendong/user/code/qwen3-tts/seed_audio/000006/female_01_zh.wav",
      "spoken_command": "向前走并抓取某物",
      "ideal_motion": "听完后开始向前走并抓取;动作尚未结束时收到下一条指令。"
    },
    {
      "turn_index": 2,
      "source_motion_filename": "wave_two_hands_R_001__A533",
      "audio_path": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/buwendong/user/code/qwen3-tts/seed_audio/000155/female_01_zh.wav",
      "spoken_command": "双手在身前挥动",
      "ideal_motion": "新语音刚出现时短暂延续旧动作,听清后平滑切入 listen,再从当时真实姿态开始挥手。"
    },
    {
      "turn_index": 3,
      "source_motion_filename": "neutral_walk_ff_270_R_001__A534_M",
      "audio_path": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/buwendong/user/code/qwen3-tts/seed_audio/000120/female_01_zh.wav",
      "spoken_command": "向左走路",
      "ideal_motion": "挥手尚未自然结束时响应新语音,随后相对当前身体朝向向左行走并最终恢复站立。"
    }
  ],
  "ideal_control_timeline": [
    "初始背景输出 </idle>,第一条语音出现后进入 </listen>",
    "每次新语音刚开始时仍执行上一轮 </motion>,不提前读取未来语义",
    "听到有效语音后由 </motion> 平滑切到 </listen>",
    "语音结束后进入新的 </motion>,最后回到 </idle>"
  ],
  "continuity_requirements": [
    "三轮共用一个 Qwen KV cache 和一段连续办公室背景",
    "打断不清空机器人状态或 RTC 历史",
    "新动作从旧动作被打断时的真实位置、朝向和关节状态接续",
    "旧动作的 future progress 仍以其原始完整时长为分母"
  ]
}