File size: 6,310 Bytes
b2e15d5 |
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 |
#!/usr/bin/env python3
"""Filter and augment Unitree G1 dataset to contain only walking fragments.
The script:
1. Reads *episode_stats.json* produced by **analyze_dataset.py**.
2. Determines whether base angular velocity appears to be in rad/s or deg/s.
3. Walks through each trajectory file, extracts contiguous segments where
linear speed <= 2 m/s, and saves them as new .pt files in an output folder.
4. Optionally performs left↔right mirroring augmentation.
Run:
python augment_dataset.py \
--root /home/ubuntu/MoCapDataset/AMASSDataset/UnitreeG1 \
--out /home/ubuntu/MoCapDataset/AMASSDataset/UnitreeG1_WalkOnly \
--mirror
"""
from __future__ import annotations
import argparse
import json
import os
from pathlib import Path
from typing import List, Dict, Tuple
import torch
import numpy as np
import isaaclab.utils.math as math_utils
_ALLOWED_EXT = {".pt", ".pth", ".pkl", ".npz"}
# ----------------------------------------------------------------------------------
# helpers
# ----------------------------------------------------------------------------------
def load_file(path: Path) -> Dict[str, torch.Tensor]:
if path.suffix in {".npz", ".pkl"}:
data = dict(np.load(path, allow_pickle=True))
for k, v in data.items():
if isinstance(v, np.ndarray) and v.dtype.kind in {"f", "c", "i", "u", "b"}:
data[k] = torch.from_numpy(v)
else:
data[k] = v
return data # type: ignore[return-value]
return torch.load(path, map_location="cpu") # type: ignore[return-value]
def save_pt(data: Dict[str, torch.Tensor], path: Path):
path.parent.mkdir(parents=True, exist_ok=True)
torch.save(data, path)
def mirror_left_right(data: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
"""Simple left↔right mirror assuming naming pattern 'left_' / 'right_'."""
mirrored = {k: v.clone() if torch.is_tensor(v) else v for k, v in data.items()}
names = data.get("joint_names", None)
if names is None:
return mirrored # cannot mirror
names = list(names)
swap = {}
for i, n in enumerate(names):
if n.startswith("left_"):
mirror_name = "right_" + n[5:]
elif n.startswith("right_"):
mirror_name = "left_" + n[6:]
else:
continue
if mirror_name in names:
swap[i] = names.index(mirror_name)
if not swap:
return mirrored
qpos = mirrored["qpos"]
qvel = mirrored["qvel"]
qpos_new = qpos.clone()
qvel_new = qvel.clone()
for i, j in swap.items():
qpos_new[:, 7+i] = qpos[:, 7+j]
qpos_new[:, 7+j] = qpos[:, 7+i]
qvel_new[:, 6+i] = qvel[:, 6+j]
qvel_new[:, 6+j] = qvel[:, 6+i]
# flip Y to mirror along sagittal plane
qpos_new[:, 1] = -qpos_new[:, 1]
qvel_new[:, 1] = -qvel_new[:, 1]
qpos_new[:, 4] = -qpos_new[:, 4] # quaternion y
qpos_new[:, 6] = -qpos_new[:, 6] # quaternion z
mirrored["qpos"] = qpos_new
mirrored["qvel"] = qvel_new
return mirrored
def extract_walking_segments(data: Dict[str, torch.Tensor], min_len: int = 50) -> List[Dict[str, torch.Tensor]]:
"""Return list of new dicts containing walking-only contiguous clips.
Filtering is done using body-frame velocities:
|vx_body| < 1.5 m/s and |vy_body| < 0.5 m/s
"""
qpos = data["qpos"]
qvel = data["qvel"]
base_lin_vel = qvel[:, :3] # (N, 3)
base_quat = qpos[:, 3:7] # (N, 4), assumed (w, x, y, z)
# Convert world-frame velocity to body-frame
base_lin_vel_body = math_utils.quat_rotate_inverse(base_quat, base_lin_vel)
# Apply thresholds
mask = (base_lin_vel_body[:, 0].abs() < 1.5) & (base_lin_vel_body[:, 1].abs() < 0.5)
segments: List[Tuple[int, int]] = []
start = None
for i, m in enumerate(mask):
if m and start is None:
start = i
elif not m and start is not None:
if i - start >= min_len:
segments.append((start, i))
start = None
# tail segment
if start is not None and len(qpos) - start >= min_len:
segments.append((start, len(qpos)))
clips = []
for s, e in segments:
clip = {
k: v[s:e].clone() if torch.is_tensor(v) and v.ndim > 0 else (v.clone() if torch.is_tensor(v) else v)
for k, v in data.items()
}
clips.append(clip)
return clips
def main():
parser = argparse.ArgumentParser(description="Filter and augment walking trajectories")
parser.add_argument("--root", required=True, type=str, help="Original dataset root")
parser.add_argument("--out", required=True, type=str, help="Output folder for walking clips")
parser.add_argument("--mirror", default=False, action="store_true", help="Generate left-right mirrored copies")
parser.add_argument("--speed_thr", type=float, default=1.5, help="Max linear speed (m/s) for walking")
parser.add_argument("--min_len", type=int, default=50, help="Minimum clip length to keep")
args = parser.parse_args()
root = Path(args.root)
out_root = Path(args.out)
out_root.mkdir(parents=True, exist_ok=True)
# iterate files
files = []
for p, _, names in os.walk(root):
for n in names:
if n == "shape_optimized.pkl":
continue
if Path(n).suffix in _ALLOWED_EXT:
files.append(Path(p) / n)
files.sort()
total_clips = 0
for f_idx, path in enumerate(files):
data = load_file(path)
clips = extract_walking_segments(data, min_len=args.min_len)
for idx, clip in enumerate(clips):
rel_dir = path.relative_to(root).parent
name = path.stem + f"_walk_{idx}.pt"
save_pt(clip, out_root / rel_dir / name)
total_clips += 1
if args.mirror:
mirror_clip = mirror_left_right(clip)
name_m = path.stem + f"_walk_{idx}_mir.pt"
save_pt(mirror_clip, out_root / rel_dir / name_m)
total_clips += 1
if (f_idx + 1) % 20 == 0:
print(f"Processed {f_idx+1}/{len(files)} files…")
print(f"Done. Saved {total_clips} walking clips to {out_root}")
if __name__ == "__main__":
main() |