You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi! Here is Tony from GRASP Lab@UPenn.
When we are evaluating the pi0, we notice current implementation of examples/droid/main.py for pi0 only records frames at policy inference intervals. This results in inconsistent video playback speeds. The delays (~750ms–1s) inherent to inference and environment steps lead to misleading visualizations that roughly appear as real-time but do not precisely reflect true action timing.
**Expected Feature: **
Videos should play back consistently at real-time speed (e.g., 10 FPS), accurately representing agent behavior.
Implement timestamp-based frame recording
Alternatively, continuously record frames independently at a fixed frame rate (e.g., client-side recording).
Hope this could help future benchmark & evaluation!
The text was updated successfully, but these errors were encountered:
Hi! Here is Tony from GRASP Lab@UPenn.
When we are evaluating the pi0, we notice current implementation of examples/droid/main.py for pi0 only records frames at policy inference intervals. This results in inconsistent video playback speeds. The delays (~750ms–1s) inherent to inference and environment steps lead to misleading visualizations that roughly appear as real-time but do not precisely reflect true action timing.
**Expected Feature: **
Videos should play back consistently at real-time speed (e.g., 10 FPS), accurately representing agent behavior.
Hope this could help future benchmark & evaluation!
The text was updated successfully, but these errors were encountered: