File size: 7,853 Bytes
319eb16
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import glob
import json
import os
import random
from pathlib import Path
from typing import Any

import cv2
import numpy as np

from dataset_upload.helpers import generate_unique_id

TASK_TO_INSTRUCTION = {
    "FailPickCube-v1": "Pick up the red cube",
    "FailPushCube-v1": "Push and move a cube to a goal region in front of it",
    "FailStackCube-v1": "Pick up a red cube and stack it on top of a green cube and let go of the cube without it falling",
}


class FailSafeFrameListLoader:
    """Pickle-able loader that reads a list of image paths on demand.

    Returns np.ndarray (T, H, W, 3) uint8.
    """

    def __init__(self, image_paths: list[str]) -> None:
        self.image_paths = image_paths
        assert len(image_paths) > 0

    def __call__(self) -> np.ndarray:
        frames: list[np.ndarray] = []
        for p in self.image_paths:
            img_bgr = cv2.imread(p, cv2.IMREAD_COLOR)
            if img_bgr is None:
                continue
            img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
            frames.append(img_rgb)
        if not frames:
            return np.empty((0, 0, 0, 3), dtype=np.uint8)
        frames_np = np.asarray(frames, dtype=np.uint8)
        return frames_np


def _sorted_pngs(dir_path: Path) -> list[str]:
    files = [str(p) for p in dir_path.glob("*.png")]

    def _key(s: str) -> tuple:
        name = os.path.splitext(os.path.basename(s))[0]
        try:
            return (int(name),)
        except Exception:
            return (name,)

    files.sort(key=_key)
    return files


def _make_traj(
    image_paths: list[str], task: str, instruction: str, is_success: bool, sub_task: str | None = None
) -> dict[str, Any]:
    traj: dict[str, Any] = {}
    traj["id"] = generate_unique_id()
    # Combine main instruction with optional sub_task for clarity
    if sub_task:
        traj["task"] = sub_task
    else:
        traj["task"] = instruction
    traj["frames"] = FailSafeFrameListLoader(image_paths)
    traj["is_robot"] = True
    traj["quality_label"] = "successful" if is_success else "failure"
    traj["data_source"] = "failsafe"
    traj["preference_group_id"] = None
    traj["preference_rank"] = None
    return traj


def _gather_full_episodes(task_dir: Path, view: str, instruction: str) -> list[dict]:
    episodes: list[dict] = []
    # Seeds are numbered directories directly under task_dir
    for seed_dir in sorted([p for p in task_dir.iterdir() if p.is_dir()]):
        # Ground truth (success)
        gt_view_dir = seed_dir / "Ground_Truth" / view
        if gt_view_dir.exists():
            imgs = _sorted_pngs(gt_view_dir)
            assert len(imgs) > 0
            if imgs:
                episodes.append(_make_traj(imgs, task_dir.name, instruction, is_success=True))

        # Failures: any subfolder except Ground_Truth
        for attempt_dir in sorted([p for p in seed_dir.iterdir() if p.is_dir() and p.name != "Ground_Truth"]):
            view_dir = attempt_dir / view
            if view_dir.exists():
                assert len(imgs) > 0
                imgs = _sorted_pngs(view_dir)
                if imgs:
                    episodes.append(_make_traj(imgs, task_dir.name, instruction, is_success=False))
    return episodes


def _gather_sub_episodes_from_json(dataset_root: Path, view: str) -> list[dict]:
    episodes: list[dict] = []
    # JSON files like vla_data_FailPickCube-v1.json, vla_data_GT_PickCube-v1.json etc.
    json_dir = dataset_root / "json_files"
    if not json_dir.exists():
        json_dir = dataset_root  # fallback if jsons are at root

    json_files = glob.glob(str(json_dir / "vla_data_*.json"))
    for jf in sorted(json_files):
        try:
            with open(jf, "r") as f:
                data = json.load(f)
        except Exception:
            continue
        if not isinstance(data, list):
            continue
        # sub sample 1/3 for 3 views
        for entry in random.sample(data, len(data) // 3):
            task_key = entry.get("task")
            instruction = entry.get("instruction") or TASK_TO_INSTRUCTION.get(task_key, task_key or "")
            sub_task = entry.get("sub_task")
            failure_type = entry.get("failure_type", "None")
            # Image list is relative to dataset root
            imgs_rel = entry.get("image", [])
            if not imgs_rel:
                continue
            # Filter by desired view: ensure each path contains "/<view>/"
            if view:
                imgs_rel = [p for p in imgs_rel if f"/{view}/" in p or f"\\{view}\\" in p]
                if len(imgs_rel) == 0:
                    continue
            image_paths = [str((dataset_root / p).resolve()) for p in imgs_rel]
            is_success = (failure_type is None) or (str(failure_type).lower() == "none")
            episodes.append(
                _make_traj(image_paths, task_key or "failsafe", instruction, is_success=is_success, sub_task=sub_task)
            )
    return episodes


def load_failsafe_dataset(dataset_path: str) -> dict[str, list[dict]]:
    """Load FailSafe dataset from local folders and JSON sub-trajectory annotations.

    Args:
        dataset_path: Root directory containing FailPickCube-v1/ FailPushCube-v1/ FailStackCube-v1/ and jsons

    Returns:
        Mapping: instruction string -> list of trajectory dicts
    """
    views = ["front", "side", "wrist"]
    include_sub_trajectories = True
    root = Path(os.path.expanduser(dataset_path))
    if not root.exists():
        raise FileNotFoundError(f"FailSafe dataset path not found: {root}")

    task_dirs = [
        p for p in [root / "FailPickCube-v1", root / "FailPushCube-v1", root / "FailStackCube-v1"] if p.exists()
    ]

    task_data: dict[str, list[dict]] = {}

    # Sub-trajectory episodes from JSON
    if include_sub_trajectories:
        for view in views:
            # sample one view
            sub_episodes = _gather_sub_episodes_from_json(root, view=view)
            print(f"Found {len(sub_episodes)} sub-trajectories for {view} after sampling 1/3 of the data")
            for traj in sub_episodes:
                task = traj["task"]
                task_data.setdefault(task, []).append(traj)

    # Full episodes
    for tdir in task_dirs:
        instruction = TASK_TO_INSTRUCTION.get(tdir.name, tdir.name)
        print(f"Gathering full episodes for {instruction}")
        for view in views:
            episodes = _gather_full_episodes(tdir, view=view, instruction=instruction)
            print(f"Found {len(episodes)} episodes for {instruction} {view}")
            if episodes:
                task_data.setdefault(instruction, []).extend(episodes)

    # only keep tasks that have both failed and successful trajectories
    task_data_paired = {}
    for task, trajectories in task_data.items():
        failed_trajectories = [t for t in trajectories if t["quality_label"] == "failure"]
        successful_trajectories = [t for t in trajectories if t["quality_label"] == "successful"]
        if len(failed_trajectories) == 0 or len(successful_trajectories) == 0:
            continue
        task_data_paired[task] = failed_trajectories + successful_trajectories

    print(
        f"Found {len(task_data_paired)} tasks with both failed and successful trajectories from originally {len(task_data)} tasks"
    )

    # print how many failed and successful trajectories there are
    failed_trajectories = [
        sum([1 for t in traj if t["quality_label"] == "failure"]) for traj in task_data_paired.values()
    ]
    successful_trajectories = [
        sum([1 for t in traj if t["quality_label"] == "successful"]) for traj in task_data_paired.values()
    ]
    print(f"Found {sum(failed_trajectories)} failed trajectories")
    print(f"Found {sum(successful_trajectories)} successful trajectories")
    return task_data_paired