## Load in Python

import pandas as pd
from pathlib import Path
import cv2

# annotations
csv_path = "annotations.csv"
df = pd.read_csv(csv_path)

# view first rows
print(df.head())

# access a clip
row = df.iloc[0]
video_file = Path(row.video_path)

# basic frame iterator
cap = cv2.VideoCapture(str(video_file))
frames = []
while True:
    ok, frame = cap.read()
    if not ok:
        break
    frames.append(frame)
cap.release()

print("frames:", len(frames))
print("object_class:", row.object_class)
print("container:", row.container_type)
print("outcome:", row.outcome)


count clips by container_type

filter outcomes to find failure clusters

group by persistence to test off-frame behavior

sample occlusion ranges for tests

does grip outcome correlate with container?

do mis-grips cluster near occlusion?

does persistence help reduce false resets?

can baseline models handle this without spatial fields?