File size: 16,113 Bytes
d62cd4c | 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 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 | """
Prepare EVAC inference inputs from RLBench-style episode data.
Input layout:
episodes_root/
βββ episode_0/
β βββ actions.npy # (T_act, 8), single-hand [xyz, quat_xyzw, gripper]
β βββ view1/
β β βββ rgb/video.mp4
β β βββ camera_params.json # {"<frame_id>": {"extrinsics":..., "intrinsics":...}}
β βββ view2/ ...
βββ episode_1/
β βββ ...
Output layout:
<user-specified output_root>/
βββ episode_0/
β βββ <view1>/ # only FIXED-camera views
β β βββ frame.png # video frame at t_start (first frame gripper visible)
β β βββ actions.npy # (T + 3, 16), dual-hand, history-padded
β β βββ extrinsics.npy # (4, 4) c2w
β β βββ intrinsics.npy # (3, 3) K (abs-valued fx, fy)
β βββ <view2>/ ...
βββ ...
Pipeline:
1. Discover view folders (contain camera_params.json).
2. Filter: keep views whose extrinsics are identical across all recorded frames.
3. Map each video frame t -> action[round(t * T_action / T_video)] (handles
non-exact ratios like 41:163 or 41:164 by clamping at the tail).
4. Find t_start: first video frame where right-hand EEF projects inside
the image with positive depth (gripper enters camera view).
5. Slice: frames [t_start, T_video), actions aligned to those frames.
6. Convert 8D single-hand -> 16D dual-hand (real on right, placeholder on left).
7. Prepend (n_previous - 1) copies of first frame to align with EVAC's history slots.
8. Write frame.png from video at t_start; write actions.npy; write K and c2w.
Usage:
python prepare_evac_input.py -i /path/to/episodes_root -o /path/to/out
python prepare_evac_input.py -i ... -o ... --hand right
python prepare_evac_input.py -i ... -o ... --fix_tol 1e-6 --n_previous 4
python prepare_evac_input.py -i ... -o ... --episodes episode_0 episode_5
"""
import argparse
import json
import os
from pathlib import Path
import cv2
import numpy as np
# ---------------------------------------------------------------------------
# Action conversion
# ---------------------------------------------------------------------------
def single_to_dual(actions_8d: np.ndarray, hand: str = "right") -> np.ndarray:
"""[T, 8] -> [T, 16]. Real data on `hand`, placeholder on the other."""
assert actions_8d.ndim == 2 and actions_8d.shape[1] == 8
T = actions_8d.shape[0]
out = np.zeros((T, 16), dtype=np.float32)
if hand == "right":
out[:, 3:7] = np.array([0, 0, 0, 1], dtype=np.float32) # identity quat
out[:, 7] = 1.0 # gripper open
out[:, 8:11] = actions_8d[:, 0:3]
out[:, 11:15] = actions_8d[:, 3:7]
out[:, 15] = actions_8d[:, 7]
elif hand == "left":
out[:, 0:3] = actions_8d[:, 0:3]
out[:, 3:7] = actions_8d[:, 3:7]
out[:, 7] = actions_8d[:, 7]
out[:, 11:15] = np.array([0, 0, 0, 1], dtype=np.float32)
out[:, 15] = 1.0
else:
raise ValueError(f"hand must be 'left' or 'right', got {hand!r}")
return out
def prepend_history_pad(actions_16d: np.ndarray, n_previous: int) -> np.ndarray:
"""Prepend (n_previous - 1) copies of the first frame."""
if n_previous <= 1:
return actions_16d
return np.concatenate([actions_16d[:1]] * (n_previous - 1) + [actions_16d], axis=0)
# ---------------------------------------------------------------------------
# Camera handling
# ---------------------------------------------------------------------------
def load_camera_params(camera_params_path: Path):
"""Parse camera_params.json -> sorted frame_ids, (T, 4, 4) ext, (T, 3, 3) K."""
with open(camera_params_path, "r") as f:
data = json.load(f)
frame_ids = sorted(data.keys())
ext = np.stack([np.array(data[k]["extrinsics"], dtype=np.float64) for k in frame_ids])
K = np.stack([np.array(data[k]["intrinsics"], dtype=np.float64) for k in frame_ids])
return frame_ids, ext, K
def is_fixed_camera(ext: np.ndarray, tol: float = 1e-6) -> bool:
"""True if extrinsics are identical across all frames (within tol)."""
if ext.shape[0] < 2:
return True
return np.abs(ext - ext[0:1]).max() < tol
def normalize_intrinsic(K_3x3: np.ndarray) -> np.ndarray:
"""Fix RLBench/OpenGL-style negative fx/fy by taking absolute values."""
K_out = K_3x3.astype(np.float32).copy()
K_out[0, 0] = abs(K_out[0, 0])
K_out[1, 1] = abs(K_out[1, 1])
return K_out
# ---------------------------------------------------------------------------
# Gripper-in-view detection via projection
# ---------------------------------------------------------------------------
def project_points_world_to_pixel(points_world: np.ndarray,
c2w: np.ndarray,
K: np.ndarray):
"""
points_world: (N, 3) world-frame 3D points
c2w: (4, 4) camera-to-world; we invert to get world-to-camera
K: (3, 3) intrinsic
Returns: (N, 2) pixel coords (u, v), (N,) camera-frame z depth
"""
w2c = np.linalg.inv(c2w)
N = points_world.shape[0]
pts_h = np.concatenate([points_world, np.ones((N, 1))], axis=1) # (N, 4)
pts_cam = (w2c @ pts_h.T).T[:, :3] # (N, 3)
uv_h = (K @ pts_cam.T).T # (N, 3)
# Avoid divide by zero / behind camera
z = uv_h[:, 2]
uv = np.zeros((N, 2), dtype=np.float64)
valid = np.abs(z) > 1e-8
uv[valid] = uv_h[valid, :2] / z[valid, None]
return uv, pts_cam[:, 2]
def find_gripper_entry_frame(eef_world_per_video_frame: np.ndarray,
c2w: np.ndarray, K: np.ndarray,
H: int, W: int,
margin: int = 0) -> int:
"""
Find first t such that EEF at video-frame t projects inside [margin, W-margin) x
[margin, H-margin) with positive depth.
eef_world_per_video_frame: (T_video, 3)
Returns index in [0, T_video), or -1 if never visible.
"""
uv, z_cam = project_points_world_to_pixel(eef_world_per_video_frame, c2w, K)
u = uv[:, 0]
v = uv[:, 1]
in_view = (
(u >= margin) & (u < W - margin) &
(v >= margin) & (v < H - margin) &
(z_cam > 0.0)
)
idx = np.where(in_view)[0]
return int(idx[0]) if len(idx) > 0 else -1
# ---------------------------------------------------------------------------
# Video I/O
# ---------------------------------------------------------------------------
def read_video_frame(video_path: Path, frame_idx: int) -> np.ndarray:
"""Return a single frame (H, W, 3) in BGR order from an .mp4 at frame_idx."""
cap = cv2.VideoCapture(str(video_path))
if not cap.isOpened():
raise IOError(f"Cannot open video: {video_path}")
cap.set(cv2.CAP_PROP_POS_FRAMES, frame_idx)
ok, frame = cap.read()
cap.release()
if not ok:
raise IOError(f"Cannot read frame {frame_idx} from {video_path}")
return frame # BGR
def video_frame_count_and_size(video_path: Path):
cap = cv2.VideoCapture(str(video_path))
if not cap.isOpened():
raise IOError(f"Cannot open video: {video_path}")
n = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
cap.release()
return n, h, w
# ---------------------------------------------------------------------------
# Action-to-video mapping
# ---------------------------------------------------------------------------
def frame_to_action_index(frame_idx: int, num_frames: int, num_actions: int) -> int:
"""
Map a video frame index to its corresponding action index.
Handles non-exact ratios (e.g. 41 video frames : 163 actions = 3.976:1)
by rounding and clamping to [0, num_actions - 1]. The last video frame
always maps to the last action, absorbing the Β±1 ragged tail.
"""
if num_frames <= 1:
return 0
# Pin the last video frame to the last action.
if frame_idx >= num_frames - 1:
return num_actions - 1
ratio = num_actions / num_frames
act_idx = int(round(frame_idx * ratio))
return max(0, min(act_idx, num_actions - 1))
def build_action_for_video_frames(actions_8d_full: np.ndarray,
n_video: int) -> np.ndarray:
"""Return (n_video, 8) by mapping each video frame idx -> action idx."""
T_act = actions_8d_full.shape[0]
idxs = np.array([frame_to_action_index(t, n_video, T_act)
for t in range(n_video)], dtype=np.int64)
return actions_8d_full[idxs]
# ---------------------------------------------------------------------------
# Per-episode per-view processing
# ---------------------------------------------------------------------------
def process_view(episode_dir: Path, view_dir: Path, out_view_dir: Path,
actions_8d_full: np.ndarray,
n_previous: int, hand: str,
fix_tol: float, margin: int, observation_offset: int,
verbose: bool = True):
"""Returns a status string for logging."""
cam_path = view_dir / "camera_params.json"
video_path = view_dir / "rgb" / "video.mp4"
if not cam_path.exists():
return f"SKIP (no camera_params.json)"
if not video_path.exists():
return f"SKIP (no rgb/video.mp4)"
# Camera params
frame_ids, ext, K_stack = load_camera_params(cam_path)
if not is_fixed_camera(ext, tol=fix_tol):
max_diff = float(np.abs(ext - ext[0:1]).max())
return f"SKIP (camera not fixed, max ext diff={max_diff:.4f})"
c2w = ext[0].astype(np.float32) # (4, 4)
K = normalize_intrinsic(K_stack[0]) # (3, 3)
# Video
n_video, H, W = video_frame_count_and_size(video_path)
if n_video < 2:
return f"SKIP (video has {n_video} frame(s))"
# Align actions to video frames via ratio mapping.
# For each video frame t in [0, n_video), pick action[round(t * T_act / n_video)].
# This handles non-exact ratios (e.g. 163:41 or 164:41) by clamping at the tail.
T_act_full = actions_8d_full.shape[0]
actions_video_rate = build_action_for_video_frames(actions_8d_full, n_video) # (n_video, 8)
# Detect gripper entry frame using right-hand EEF xyz (cols 0:3 in the 8D layout).
eef_world_seq = actions_video_rate[:, 0:3].astype(np.float32) # (n_video, 3)
t_entry = find_gripper_entry_frame(eef_world_seq, c2w, K,
H=H, W=W, margin=margin)
if t_entry < 0:
return f"SKIP (gripper never projects into view; video={n_video})"
# Apply observation offset: start a few frames AFTER entry so the gripper
# is meaningfully inside the frame, not just clipping the edge.
t_start = t_entry + observation_offset
if t_start >= n_video - 1:
return (f"SKIP (t_start={t_start} (entry={t_entry} + offset={observation_offset}) "
f">= n_video={n_video})")
# Slice from t_start
actions_sliced_8d = actions_video_rate[t_start:] # (T_out, 8)
T_out = actions_sliced_8d.shape[0]
# Convert to dual-hand + history pad
a16 = single_to_dual(actions_sliced_8d, hand=hand)
a16 = prepend_history_pad(a16, n_previous=n_previous) # (T_out + 3, 16)
# Grab video frame at t_start (BGR); save as PNG
frame_bgr = read_video_frame(video_path, t_start)
# Write outputs
out_view_dir.mkdir(parents=True, exist_ok=True)
cv2.imwrite(str(out_view_dir / "frame.png"), frame_bgr)
np.save(out_view_dir / "actions.npy", a16)
np.save(out_view_dir / "extrinsics.npy", c2w)
np.save(out_view_dir / "intrinsics.npy", K)
return (f"OK (entry={t_entry}, t_start={t_start}, T_out={T_out}, padded={a16.shape[0]}, "
f"video={n_video}x{H}x{W}, actions={T_act_full}, ratio={T_act_full/n_video:.3f})")
# ---------------------------------------------------------------------------
# Top-level walker
# ---------------------------------------------------------------------------
def process_episode(ep_dir: Path, out_ep_dir: Path,
n_previous: int, hand: str,
fix_tol: float, margin: int, observation_offset: int,
verbose: bool = True):
actions_path = ep_dir / "actions.npy"
if not actions_path.exists():
print(f"[{ep_dir.name}] SKIP: no actions.npy")
return
actions_8d_full = np.load(actions_path)
if actions_8d_full.ndim != 2 or actions_8d_full.shape[1] != 8:
print(f"[{ep_dir.name}] SKIP: actions.npy shape {actions_8d_full.shape} != (T, 8)")
return
view_dirs = [d for d in sorted(ep_dir.iterdir())
if d.is_dir() and (d / "camera_params.json").exists()]
if not view_dirs:
print(f"[{ep_dir.name}] SKIP: no view folders with camera_params.json")
return
for view_dir in view_dirs:
out_view_dir = out_ep_dir / view_dir.name
status = process_view(
episode_dir=ep_dir, view_dir=view_dir, out_view_dir=out_view_dir,
actions_8d_full=actions_8d_full,
n_previous=n_previous, hand=hand,
fix_tol=fix_tol, margin=margin,
observation_offset=observation_offset,
verbose=verbose,
)
print(f"[{ep_dir.name}/{view_dir.name}] {status}")
def main():
p = argparse.ArgumentParser(
formatter_class=argparse.RawDescriptionHelpFormatter, description=__doc__
)
p.add_argument("-i", "--input_root", required=True, type=Path,
help="Root folder containing episode_* subfolders")
p.add_argument("-o", "--output_root", required=True, type=Path,
help="Output folder (will be created)")
p.add_argument("--episodes", nargs="*", default=None,
help="Only process these episode subfolder names (default: all)")
p.add_argument("--n_previous", type=int, default=4,
help="EVAC history length to pad (default 4)")
p.add_argument("--hand", choices=["left", "right"], default="right",
help="Which side of the 16D layout gets the real data")
p.add_argument("--fix_tol", type=float, default=1e-6,
help="Max per-element extrinsics diff to count as 'fixed camera'")
p.add_argument("--margin", type=int, default=0,
help="Pixel margin for gripper-in-view check (default 0)")
p.add_argument("--observation_offset", type=int, default=3,
help="Number of video frames to advance past the gripper-entry "
"frame before taking observation (default 2)")
args = p.parse_args()
if not args.input_root.exists():
raise FileNotFoundError(args.input_root)
args.output_root.mkdir(parents=True, exist_ok=True)
# Discover episode folders
ep_dirs = sorted([d for d in args.input_root.iterdir() if d.is_dir()])
if args.episodes:
wanted = set(args.episodes)
ep_dirs = [d for d in ep_dirs if d.name in wanted]
print(f"Found {len(ep_dirs)} episode(s) to process in {args.input_root}")
print(f"Output -> {args.output_root}")
print(f"Params: n_previous={args.n_previous}, hand={args.hand}, "
f"fix_tol={args.fix_tol}, margin={args.margin}, "
f"observation_offset={args.observation_offset}")
print("-" * 70)
for ep_dir in ep_dirs:
out_ep_dir = args.output_root / ep_dir.name
process_episode(
ep_dir=ep_dir, out_ep_dir=out_ep_dir,
n_previous=args.n_previous,
hand=args.hand, fix_tol=args.fix_tol, margin=args.margin,
observation_offset=args.observation_offset,
)
if __name__ == "__main__":
main() |