from __future__ import annotations AUDIO_RATE = 16_000 MOTION_FPS = 50 AUDIO_WINDOW_SECONDS = 0.6 # Qwen consumes one overlapping audio window every 0.4 seconds. AUDIO_STRIDE_FRAMES = 20 # Dense action supervision shifts the current state by one 50 Hz frame. ACTION_SAMPLE_STEP_FRAMES = 1 ACTION_CHUNK_FRAMES = 30 ACTION_EXEC_FRAMES = 20 RTC_PREFIX_FRAMES = 10 def complete_audio_clip_starts( total_samples: int, *, sampling_rate: int = AUDIO_RATE, window_seconds: float = AUDIO_WINDOW_SECONDS, stride_frames: int = AUDIO_STRIDE_FRAMES, motion_fps: int = MOTION_FPS, ) -> list[int]: """Return the exact complete-window starts used by the training exporter.""" window_samples = int(round(float(window_seconds) * int(sampling_rate))) stride_samples = int(round(int(stride_frames) / float(motion_fps) * int(sampling_rate))) if total_samples < window_samples: return [] return list(range(0, int(total_samples) - window_samples + 1, stride_samples)) def aligned_motion_start_frame( total_samples: int, *, sampling_rate: int = AUDIO_RATE, motion_fps: int = MOTION_FPS, ) -> int: """First motion frame at or after the exact decoded audio end sample.""" return (int(total_samples) * int(motion_fps) + int(sampling_rate) - 1) // int( sampling_rate )