Robotics
behavior-1k
pi0.5
vla
File size: 886 Bytes
5ab0bd2
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
diff --git a/OmniGibson/omnigibson/learning/utils/lerobot_utils.py b/OmniGibson/omnigibson/learning/utils/lerobot_utils.py
index 95eb5e1b5..f69c6a13d 100644
--- a/OmniGibson/omnigibson/learning/utils/lerobot_utils.py
+++ b/OmniGibson/omnigibson/learning/utils/lerobot_utils.py
@@ -92,7 +92,8 @@ def decode_video_frames_torchvision(
 
     # set the first and last requested timestamps
     # Note: previous timestamps are usually loaded, since we need to access the previous key frame
-    first_ts = min(timestamps) - 5  # a little backward to account for timestamp mismatch
+    # Videos were re-encoded with a keyframe every 4 frames, so a short backoff is enough (was 5s = ~150 decoded frames).
+    first_ts = min(timestamps) - 0.5  # a little backward to account for timestamp mismatch
     last_ts = max(timestamps)
 
     # access closest key frame of the first requested frame