File size: 4,972 Bytes
f91b300
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Detectors for unusable frames, and the clean-span complement.

Three failure modes are flagged per episode:

``intensity_spikes``   a GelSight reading far above anything contact produces —
                       usually the sensor being knocked or re-seated
``pose_teleports_*``   OptiTrack solving to the wrong marker set, which moves
                       the sensor implausibly far *and* rotates it implausibly
                       fast in a single frame
``ot_loss_*``          the tracker dropping out, which shows up as a run of
                       bit-identical poses rather than as missing samples

Thresholds are the ones validated against the published motherboard
``bad_frames.json`` (25/27 episodes bit-identical).

Previously split across ``detect_bad_intervals.py`` and ``build_segments.py``
in ``twm/scripts/``; the latter has since been archived, so this module is now
the only live copy of ``find_clean_segments``.
"""
from __future__ import annotations

import numpy as np

from .config import FPS

TAU_INTENSITY = 30.0
TAU_VELOCITY_MPS = 5.0
TAU_ANGULAR_RAD_PS = 15.0
FREEZE_THRESHOLD_S = 0.25
BUFFER_FRAMES = 3
EPS_POSE_BIT = 1e-7


def merge_intervals(events, gap: int = 1) -> list[list[int]]:
    """Merge inclusive ``(a, b)`` intervals that touch or overlap."""
    if not events:
        return []
    ordered = sorted((int(a), int(b)) for a, b in events)
    merged = [list(ordered[0])]
    for a, b in ordered[1:]:
        if a <= merged[-1][1] + gap:
            merged[-1][1] = max(merged[-1][1], b)
        else:
            merged.append([a, b])
    return merged


def pad_and_merge(events, T: int, buffer: int) -> list[list[int]]:
    """Pad each interval by ``±buffer``, clip to ``[0, T-1]``, then merge."""
    if not events:
        return []
    return merge_intervals([(max(0, a - buffer), min(T - 1, b + buffer))
                            for a, b in events])


def detect_intensity_spikes(intens_l: np.ndarray, intens_r: np.ndarray,
                            T: int) -> list[list[int]]:
    """Frames where either sensor reads above ``TAU_INTENSITY``."""
    above = (intens_l > TAU_INTENSITY) | (intens_r > TAU_INTENSITY)
    return pad_and_merge([(int(i), int(i)) for i in np.where(above)[0]],
                         T, BUFFER_FRAMES)


def detect_pose_teleports(pose: np.ndarray, T: int) -> list[list[int]]:
    """Frames whose pose jump is implausible in translation *and* rotation.

    The conjunction matters: ordinary fast motion trips the translational
    threshold on its own, so requiring both is what separates a tracking error
    from a quick reach.
    """
    if T < 2:
        return []
    xyz, quat = pose[:, :3], pose[:, 3:]
    qn = quat / np.maximum(np.linalg.norm(quat, axis=1, keepdims=True), 1e-12)
    trans_vel = np.linalg.norm(np.diff(xyz, axis=0), axis=1) * FPS
    dot = np.abs((qn[:-1] * qn[1:]).sum(axis=1)).clip(-1.0, 1.0)
    ang_vel = 2.0 * np.arccos(dot) * FPS
    flag = (trans_vel > TAU_VELOCITY_MPS) & (ang_vel > TAU_ANGULAR_RAD_PS)
    return pad_and_merge([(int(i), int(i + 1)) for i in np.where(flag)[0]],
                         T, BUFFER_FRAMES)


def detect_pose_freezes(pose: np.ndarray, T: int) -> list[list[int]]:
    """Runs of bit-identical pose lasting at least ``FREEZE_THRESHOLD_S``.

    OptiTrack repeats its last solution when it loses the marker set, so a
    frozen pose is track loss rather than genuine stillness — a real hold still
    jitters in the last decimal places.

    Reported unpadded, matching the published ``bad_frames.json``.
    """
    if T < 2:
        return []
    same = np.zeros(T, dtype=bool)
    same[1:] = np.all(np.abs(np.diff(pose, axis=0)) < EPS_POSE_BIT, axis=1)
    min_frames = int(round(FREEZE_THRESHOLD_S * FPS))

    events, i = [], 1
    while i < T:
        if not same[i]:
            i += 1
            continue
        j = i
        while j < T and same[j]:
            j += 1
        # the run includes the anchor frame at i-1 that the copies match
        run_a, run_b = i - 1, j - 1
        if (run_b - run_a + 1) >= min_frames:
            events.append((run_a, run_b))
        i = j
    return pad_and_merge(events, T, 0)


def find_clean_segments(T: int, bad_intervals) -> list[tuple[int, int]]:
    """Complement of the bad intervals: inclusive ``[a, b]`` clean spans."""
    segments, prev_end = [], -1
    for a, b in merge_intervals(bad_intervals):
        if a > prev_end + 1:
            segments.append((prev_end + 1, a - 1))
        prev_end = max(prev_end, b)
    if prev_end < T - 1:
        segments.append((prev_end + 1, T - 1))
    return segments


def thresholds() -> dict:
    """The detector settings, for recording alongside the results."""
    return {
        "tau_intensity": TAU_INTENSITY,
        "tau_velocity_mps": TAU_VELOCITY_MPS,
        "tau_angular_rad_per_s": TAU_ANGULAR_RAD_PS,
        "freeze_threshold_s": FREEZE_THRESHOLD_S,
        "buffer_frames": BUFFER_FRAMES,
    }