File size: 19,388 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 | #!/usr/bin/env python3
"""Dataset analysis utility for the Unitree G1 AMASS trajectories.
Run:
python analyze_dataset.py --root /home/ubuntu/MoCapDataset/AMASSDataset/UnitreeG1
The script outputs two files in the same folder:
├─ episode_stats.json – per-trajectory statistics
└─ aggregate_stats.json – averages / extrema over the whole dataset
Useful for diagnosing AMP training issues (range mismatch, outliers, etc.).
"""
from __future__ import annotations
import argparse
import json
import os
from pathlib import Path
from collections import defaultdict
from typing import Any, Dict, List
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]:
"""Load torch / numpy trajectory file into a dict of torch tensors."""
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 # keep python objects (e.g. list[str]) as-is
return data # type: ignore[return-value]
# torch files
return torch.load(path, map_location="cpu") # type: ignore[return-value]
def _tensor_range(t: torch.Tensor):
return t.min().item(), t.max().item()
def _analyze_field(value: Any) -> Dict[str, Any]:
"""Analyze a field value and return metadata about it."""
if isinstance(value, torch.Tensor):
return {
"type": "tensor",
"dtype": str(value.dtype),
"shape": list(value.shape),
"ndim": value.ndim,
"size": value.numel(),
}
elif isinstance(value, np.ndarray):
return {
"type": "numpy_array",
"dtype": str(value.dtype),
"shape": list(value.shape),
"ndim": value.ndim,
"size": value.size,
}
elif isinstance(value, list):
return {
"type": "list",
"length": len(value),
"element_type": type(value[0]).__name__ if value else "unknown",
}
elif isinstance(value, (str, int, float, bool)):
return {
"type": type(value).__name__,
"value": value if isinstance(value, (int, float, bool)) else f"<string of length {len(value)}>",
}
else:
return {
"type": type(value).__name__,
"repr": str(value)[:100] + ("..." if len(str(value)) > 100 else ""),
}
def _determine_angular_unit(max_vel: float) -> str:
"""Determine if angular velocity is in deg/s or rad/s based on magnitude."""
return "deg/s" if max_vel > 20.0 else "rad/s"
def _analyze_angular_velocity(velocities: torch.Tensor, name: str = "") -> Dict[str, Any]:
"""Analyze angular velocity data to determine if it's in deg/s or rad/s.
Uses multiple heuristics:
1. Physical limits - robots rarely exceed 1000 deg/s or ~17 rad/s
2. Distribution of values - deg/s values tend to be larger
3. Common ranges for motion capture data
Args:
velocities: Tensor of angular velocities
name: Name of the joint/axis for reporting
Returns:
Dict with analysis results
"""
abs_max = float(torch.abs(velocities).max())
abs_mean = float(torch.abs(velocities).mean())
# Convert to both units for analysis
if abs_max > 20.0: # Assuming it might be deg/s
rad_max = abs_max * np.pi / 180
rad_mean = abs_mean * np.pi / 180
deg_max = abs_max
deg_mean = abs_mean
original_unit = "deg/s"
else: # Assuming it might be rad/s
rad_max = abs_max
rad_mean = abs_mean
deg_max = abs_max * 180 / np.pi
deg_mean = abs_mean * 180 / np.pi
original_unit = "rad/s"
# Scoring system for unit determination
deg_score = 0
rad_score = 0
# Physical limits check
if deg_max > 1000: # Unusually high for deg/s
rad_score += 2
if rad_max > 17: # Unusually high for rad/s (~1000 deg/s)
deg_score += 2
# Common ranges check
if 20 <= deg_max <= 720: # Common range for deg/s in motion capture
deg_score += 1
if 0.3 <= rad_max <= 12: # Common range for rad/s in motion capture
rad_score += 1
# Mean value check
if 5 <= deg_mean <= 180: # Common mean range for deg/s
deg_score += 1
if 0.1 <= rad_mean <= 3: # Common mean range for rad/s
rad_score += 1
# Determine most likely unit
likely_unit = "deg/s" if deg_score > rad_score else "rad/s"
confidence = abs(deg_score - rad_score) / (deg_score + rad_score) if (deg_score + rad_score) > 0 else 0
return {
"likely_unit": likely_unit,
"confidence": confidence,
"max_value": abs_max,
"mean_value": abs_mean,
"deg_score": deg_score,
"rad_score": rad_score,
"analysis": {
"deg/s": {"max": deg_max, "mean": deg_mean},
"rad/s": {"max": rad_max, "mean": rad_mean}
},
"original_unit": original_unit
}
# ----------------------------------------------------------------------------------
# main analysis
# ----------------------------------------------------------------------------------
def analyse_dataset(root: Path) -> None:
# discover files
files: List[Path] = []
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()
if not files:
raise RuntimeError(f"No trajectory files found under {root}")
print(f"Found {len(files)} trajectory files. Analysing…")
# aggregate accumulators
agg = {
"num_episodes": len(files),
"total_frames": 0,
"lengths": [],
"base_lin_vel_b": {"min": torch.full((3,), torch.inf), "max": torch.full((3,), -torch.inf)},
"base_ang_vel_b": {"min": torch.full((3,), torch.inf), "max": torch.full((3,), -torch.inf)},
"base_height": {"min": torch.tensor(torch.inf), "max": torch.tensor(-torch.inf)},
"base_quat": {"min": torch.full((4,), torch.inf), "max": torch.full((4,), -torch.inf)},
"joint_pos": {}, # type: ignore[dict[str, Any]]
"joint_vel": {}, # type: ignore[dict[str, Any]]
"base_pos": {"min": torch.full((3,), torch.inf), "max": torch.full((3,), -torch.inf)},
"start_pos_list": [],
}
# field analysis accumulator
field_analysis = defaultdict(lambda: {
"count": 0,
"files_with_field": [],
"metadata": None,
"consistent_shape": True,
"shapes_seen": set(),
})
episode_stats: List[Dict[str, Any]] = []
for f_idx, path in enumerate(files):
data = _load_file(path)
# Analyze all fields in this file
for field_name, field_value in data.items():
field_info = field_analysis[field_name]
field_info["count"] += 1
field_info["files_with_field"].append(str(path.relative_to(root)))
# Analyze field metadata
metadata = _analyze_field(field_value)
if field_info["metadata"] is None:
field_info["metadata"] = metadata
# Track shape consistency for tensors/arrays
if metadata["type"] in ["tensor", "numpy_array"]:
shape_tuple = tuple(metadata["shape"])
field_info["shapes_seen"].add(shape_tuple)
if len(field_info["shapes_seen"]) > 1:
field_info["consistent_shape"] = False
# Continue with existing analysis for specific fields
if "qpos" not in data or "qvel" not in data:
print(f"[WARN] {path.name}: Missing qpos or qvel, skipping detailed analysis")
continue
qpos, qvel = data["qpos"].float(), data["qvel"].float()
n = qpos.shape[0]
agg["total_frames"] += n
agg["lengths"].append(n)
# root quantities
base_pos = qpos[:, :3]
base_quat = math_utils.quat_unique(qpos[:, 3:7]) # shape [n,4]
base_lin_vel = qvel[:, :3]
base_ang_vel = qvel[:, 3:6]
base_lin_vel_b = math_utils.quat_rotate_inverse(base_quat, base_lin_vel)
base_ang_vel_b = math_utils.quat_rotate_inverse(base_quat, base_ang_vel)
height = base_pos[:, 2]
# update global min/max
agg["base_height"]["min"] = torch.minimum(agg["base_height"]["min"], height.min())
agg["base_height"]["max"] = torch.maximum(agg["base_height"]["max"], height.max())
for k, tensor in zip(["base_lin_vel_b", "base_ang_vel_b"], [base_lin_vel_b, base_ang_vel_b]):
agg[k]["min"] = torch.minimum(agg[k]["min"], tensor.min(dim=0).values)
agg[k]["max"] = torch.maximum(agg[k]["max"], tensor.max(dim=0).values)
agg["base_quat"]["min"] = torch.minimum(agg["base_quat"]["min"], base_quat.min(dim=0).values)
agg["base_quat"]["max"] = torch.maximum(agg["base_quat"]["max"], base_quat.max(dim=0).values)
# update base_pos ranges
agg["base_pos"]["min"] = torch.minimum(agg["base_pos"]["min"], base_pos.min(dim=0).values)
agg["base_pos"]["max"] = torch.maximum(agg["base_pos"]["max"], base_pos.max(dim=0).values)
# save starting position
agg["start_pos_list"].append(base_pos[0].tolist())
# joint ranges per episode
joint_pos = qpos[:, 7:]
joint_vel = qvel[:, 6:] # skip floating base velocities (first 6)
num_joints = joint_pos.shape[1]
joint_names = data.get("joint_names", None)
if joint_names is not None and len(joint_names) == num_joints + 1:
# Likely the first entry is the floating base/root joint which is not in joint_pos
joint_names = joint_names[1:]
if joint_names is None or len(joint_names) != num_joints:
# Generate fallback names if missing or mismatched
if joint_names is not None and len(joint_names) != num_joints:
print(f"[WARN] {path.name}: joint_names length {len(joint_names)} != joint dim {num_joints}. Using generic names.")
joint_names = [f"joint_{i}" for i in range(num_joints)]
ep_joint_range = {}
ep_joint_vel_range = {}
for j in range(num_joints):
name = joint_names[j]
# Position ranges
j_min, j_max = _tensor_range(joint_pos[:, j])
ep_joint_range[name] = {"min": j_min, "max": j_max}
# Velocity ranges
v_min, v_max = _tensor_range(joint_vel[:, j])
ep_joint_vel_range[name] = {"min": v_min, "max": v_max}
# accumulate global position ranges
if name not in agg["joint_pos"]:
agg["joint_pos"][name] = {"min": j_min, "max": j_max}
else:
agg["joint_pos"][name]["min"] = min(agg["joint_pos"][name]["min"], j_min)
agg["joint_pos"][name]["max"] = max(agg["joint_pos"][name]["max"], j_max)
# accumulate global velocity ranges
if name not in agg["joint_vel"]:
agg["joint_vel"][name] = {"min": v_min, "max": v_max}
else:
agg["joint_vel"][name]["min"] = min(agg["joint_vel"][name]["min"], v_min)
agg["joint_vel"][name]["max"] = max(agg["joint_vel"][name]["max"], v_max)
# Analyze base angular velocities
base_ang_vel_analysis = {}
for i, axis in enumerate(['x', 'y', 'z']):
base_ang_vel_axis = base_ang_vel_b[:, i]
base_ang_vel_analysis[axis] = _analyze_angular_velocity(base_ang_vel_axis, f"base_ang_vel_{axis}")
# Analyze joint velocities
joint_vel_analysis = {}
for j in range(num_joints):
name = joint_names[j]
joint_vel_analysis[name] = _analyze_angular_velocity(joint_vel[:, j], name)
# store per-episode stats
episode_stats.append({
"file": str(path.relative_to(root)),
"length": n,
"base_height": _tensor_range(height),
"base_pos_start": base_pos[0].tolist(),
"base_pos_range": {
"min": base_pos.min(dim=0).values.tolist(),
"max": base_pos.max(dim=0).values.tolist(),
},
"base_lin_vel_b": {
"min": base_lin_vel_b.min(dim=0).values.tolist(),
"max": base_lin_vel_b.max(dim=0).values.tolist(),
},
"base_ang_vel_b": {
"min": base_ang_vel_b.min(dim=0).values.tolist(),
"max": base_ang_vel_b.max(dim=0).values.tolist(),
},
"base_quat": {k: v.tolist() for k, v in agg["base_quat"].items()},
"joint_pos_range": ep_joint_range,
"joint_vel_range": ep_joint_vel_range,
"base_ang_vel_analysis": base_ang_vel_analysis,
"joint_vel_analysis": joint_vel_analysis,
})
if (f_idx + 1) % 50 == 0:
print(f"Processed {f_idx+1}/{len(files)} files…")
# Process field analysis results
# print("\n" + "="*60)
# print("DATASET FIELD ANALYSIS")
# print("="*60)
# all_fields = sorted(field_analysis.keys())
# for field_name in all_fields:
# info = field_analysis[field_name]
# print(f"\nField: '{field_name}'")
# print(f" Present in: {info['count']}/{len(files)} files ({info['count']/len(files)*100:.1f}%)")
# if info['metadata']:
# meta = info['metadata']
# if meta['type'] in ['tensor', 'numpy_array']:
# if info['consistent_shape']:
# print(f" Type: {meta['type']} ({meta['dtype']})")
# print(f" Shape: {meta['shape']}")
# else:
# print(f" Type: {meta['type']} ({meta['dtype']}) - INCONSISTENT SHAPES")
# print(f" Shapes seen: {sorted(info['shapes_seen'])}")
# elif meta['type'] == 'list':
# print(f" Type: {meta['type']} (length: {meta['length']}, elements: {meta['element_type']})")
# else:
# print(f" Type: {meta['type']}")
# if 'value' in meta:
# print(f" Value: {meta['value']}")
# Create field summary for JSON output
field_summary = {}
for field_name, info in field_analysis.items():
field_summary[field_name] = {
"present_in_files": info["count"],
"present_in_percentage": round(info["count"] / len(files) * 100, 1),
"metadata": info["metadata"],
"consistent_shape": info["consistent_shape"],
}
if not info["consistent_shape"]:
field_summary[field_name]["shapes_seen"] = sorted([list(s) for s in info["shapes_seen"]])
# final aggregate statistics
agg_stats = {
"num_episodes": agg["num_episodes"],
"average_length": float(np.mean(agg["lengths"])),
"min_length": int(min(agg["lengths"])),
"max_length": int(max(agg["lengths"])),
"base_height": {k: v.item() if torch.is_tensor(v) else v.tolist() for k, v in agg["base_height"].items()},
"base_lin_vel_b": {k: v.tolist() for k, v in agg["base_lin_vel_b"].items()},
"base_ang_vel_b": {k: v.tolist() for k, v in agg["base_ang_vel_b"].items()},
"base_quat": {k: v.tolist() for k, v in agg["base_quat"].items()},
"joint_pos_global_range": agg["joint_pos"],
"joint_vel_global_range": agg["joint_vel"], # Added joint velocity global ranges
"base_pos": {k: v.tolist() for k, v in agg["base_pos"].items()},
"avg_start_pos": np.mean(agg["start_pos_list"], axis=0).tolist(),
"field_analysis": field_summary,
"angular_velocity_analysis": {
"base": base_ang_vel_analysis,
"joints": joint_vel_analysis,
}
}
# Determine angular-velocity unit (deg/s vs rad/s)
max_ang_vel = max(abs(x) for x in agg_stats["base_ang_vel_b"]["max"])
base_ang_unit = _determine_angular_unit(max_ang_vel)
agg_stats["base_ang_vel_unit"] = base_ang_unit
# Analyze joint velocity units
joint_vel_units = {}
max_joint_vels = {}
for joint_name, vel_range in agg["joint_vel"].items():
max_vel = max(abs(vel_range["min"]), abs(vel_range["max"]))
max_joint_vels[joint_name] = max_vel
joint_vel_units[joint_name] = _determine_angular_unit(max_vel)
# Check if all joints use the same unit
unique_units = set(joint_vel_units.values())
if len(unique_units) == 1:
joint_vel_unit = next(iter(unique_units))
print(f"\nAll joint velocities appear to be in {joint_vel_unit}")
else:
print("\nWARNING: Inconsistent joint velocity units detected:")
for unit in unique_units:
joints = [name for name, u in joint_vel_units.items() if u == unit]
print(f" {unit}: {', '.join(joints)}")
# Add joint velocity analysis to aggregate stats
agg_stats["joint_vel_units"] = joint_vel_units
agg_stats["joint_vel_max_magnitude"] = max_joint_vels
print(f"\nEstimated base angular velocity unit: {base_ang_unit} (max ω ≈ {max_ang_vel:.2f})")
print("Joint velocity analysis added to aggregate_stats.json")
# Print detailed analysis
print("\nAngular Velocity Analysis:")
print("\nBase Angular Velocity:")
for axis, analysis in base_ang_vel_analysis.items():
print(f" {axis}-axis: Likely {analysis['likely_unit']} (confidence: {analysis['confidence']:.2f})")
print(f" Max: {analysis['max_value']:.2f} {analysis['original_unit']}")
print(f" Mean: {analysis['mean_value']:.2f} {analysis['original_unit']}")
print("\nJoint Velocities:")
for joint, analysis in joint_vel_analysis.items():
print(f" {joint}: Likely {analysis['likely_unit']} (confidence: {analysis['confidence']:.2f})")
print(f" Max: {analysis['max_value']:.2f} {analysis['original_unit']}")
print(f" Mean: {analysis['mean_value']:.2f} {analysis['original_unit']}")
# write files
(root / "episode_stats.json").write_text(json.dumps(episode_stats, indent=2))
(root / "aggregate_stats.json").write_text(json.dumps(agg_stats, indent=2))
print(f"\nAnalysis complete. Results saved to episode_stats.json and aggregate_stats.json")
# print(f"Field analysis included for {len(all_fields)} unique fields found across all files.")
# ----------------------------------------------------------------------------------
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Analyse Unitree G1 AMASS dataset")
parser.add_argument("--root", type=str, required=True, help="Root folder of trajectories")
args = parser.parse_args()
analyse_dataset(Path(args.root).expanduser())
|