"""Extract action JSONs for the action-conditioned video2world model. Auto-detects the dataset's action layout from `meta/modality.json` and writes JSONs in the schema expected by `video2world_action.get_action_sequence`: { "action": (T, 6) -- per-frame 6D delta, pre-divided by 20.0 (inference re-multiplies by 20) "continuous_gripper_state": (T+1,) -- gripper open/close ratio in [0, 1] } Two source formats are supported: (a) GR00T 1027-D layout (e.g. In-lab_Eval) -- has `delta_right_wrist` block of shape (T, 144) = 24 future-frames * 6D. We take the first 6D as the next-frame delta. This matches training distribution. (b) GR00T 52-D layout (e.g. EgoDex_Eval, DreamDojo-HV_Eval) -- only has absolute joint angles. We compute joint-space delta via `np.diff` on right_arm[:, :6]. This is best-effort: the model was trained on EE-space deltas, so generation quality will be lower than (a). """ import argparse import glob import json import os import re import numpy as np import pandas as pd from tqdm import tqdm INFERENCE_SCALE = 20.0 def parse_args(): parser = argparse.ArgumentParser() parser.add_argument("--dataset_path", type=str, required=True, help="Path to a leaf LeRobot dataset (containing meta/, data/, videos/)") parser.add_argument("--save_path", type=str, required=True, help="Where to save extracted action JSONs") parser.add_argument("--output_json", type=str, required=True, help="Path to the batch_input.json") parser.add_argument("--output_video_dir", type=str, default="./output/action_gen", help="Directory used in the batch_input.json's output_video field") parser.add_argument("--video_subdir", type=str, default="observation.images.ego_view_freq20", help="Video subdirectory under videos/chunk-*/") parser.add_argument("--stride", type=int, default=5, help="Subsample stride to convert source FPS -> model FPS " "(default 5 = 20Hz dataset -> 4Hz model). Each output " "delta then represents motion over 1/4 s, matching training.") return parser.parse_args() def load_modality(dataset_path): path = os.path.join(dataset_path, "meta", "modality.json") with open(path) as f: return json.load(f) def extract_action_ee(raw_actions, modality, stride): """Return (T_out, 6) array of per-frame deltas, pre-divided by INFERENCE_SCALE. `stride` subsamples the source-FPS actions to the model FPS so each delta represents motion over one model-frame (e.g. 20Hz -> 4Hz with stride=5). """ action_def = modality["action"] if "delta_right_wrist" in action_def: s = action_def["delta_right_wrist"]["start"] # The 144-D block is 24 future-frame deltas at source FPS. # First 6D = +1-frame delta. To get a +stride-frame delta at model FPS, # use the slot at index (stride-1)*6. slot = (stride - 1) * 6 delta = raw_actions[::stride, s + slot:s + slot + 6] return delta / INFERENCE_SCALE, f"delta_right_wrist[slot={slot}], stride={stride}" if "right_arm" in action_def: s = action_def["right_arm"]["start"] arm = raw_actions[:, s:s + 6] # Stride-spaced finite difference: delta[i] = arm[(i+1)*stride] - arm[i*stride] arm_ds = arm[::stride] # (T_out+1, 6) ideally if arm_ds.shape[0] < 2: raise ValueError("Episode too short for stride") delta = np.diff(arm_ds, axis=0, prepend=arm_ds[:1]) # (T_out, 6) return delta / INFERENCE_SCALE, f"right_arm_diff, stride={stride} (best-effort)" raise ValueError("No usable action channel found in modality.json") def extract_gripper(raw_actions, modality, stride, length): """Return (length,) gripper state in [0, 1] aligned to the strided actions.""" action_def = modality["action"] if "right_hand" not in action_def: return np.zeros(length, dtype=np.float32) s = action_def["right_hand"]["start"] g = raw_actions[::stride, s].astype(np.float32)[:length] if g.shape[0] < length: g = np.pad(g, (0, length - g.shape[0]), mode="edge") g_min, g_max = float(g.min()), float(g.max()) if g_max - g_min > 1e-6: g = (g - g_min) / (g_max - g_min) else: g = np.zeros_like(g) return g def main(): args = parse_args() modality = load_modality(args.dataset_path) video_files = sorted(glob.glob( os.path.join(args.dataset_path, "videos", "**", "episode_*.mp4"), recursive=True, )) # Optionally filter by camera subdir. if args.video_subdir: video_files = [v for v in video_files if args.video_subdir in v] os.makedirs(args.save_path, exist_ok=True) batch_input = [] print(f"[extract] dataset={args.dataset_path}") print(f"[extract] found {len(video_files)} videos") fmt_used = None for vid_path in tqdm(video_files): match = re.search(r"episode_(\d+)", os.path.basename(vid_path)) if not match: continue ep_idx = int(match.group(1)) parquets = glob.glob( os.path.join(args.dataset_path, "data", "**", f"episode_{ep_idx:06d}.parquet"), recursive=True, ) if not parquets: print(f"[warn] parquet for episode {ep_idx} not found") continue df = pd.read_parquet(parquets[0]) raw_actions = np.stack(df["action"].values) # (T, D) action_ee, fmt = extract_action_ee(raw_actions, modality, args.stride) gripper = extract_gripper(raw_actions, modality, args.stride, action_ee.shape[0]) fmt_used = fmt # video2world_action.get_action_sequence does gripper[1:], so we pad to T+1. gripper_extended = np.append(gripper, gripper[-1]) anno = { "action": action_ee.tolist(), "continuous_gripper_state": gripper_extended.tolist(), } anno_path = os.path.join(args.save_path, f"episode_{ep_idx:06d}_action.json") with open(anno_path, "w") as f: json.dump(anno, f) batch_input.append({ "input_video": vid_path, "input_annotation": anno_path, "output_video": os.path.join(args.output_video_dir, f"gen_{os.path.basename(vid_path)}"), }) os.makedirs(os.path.dirname(args.output_json), exist_ok=True) with open(args.output_json, "w") as f: json.dump(batch_input, f, indent=4) print(f"[extract] action format: {fmt_used}") print(f"[extract] wrote {len(batch_input)} entries -> {args.output_json}") if __name__ == "__main__": main()