React Dataset · Latency Correction

Does +15f fix it?

Each clip is a vertical 2-up of the SAME moment. 20 s @ 1×, sampled at early / mid / late positions of 2 episodes per task — to check the correction holds across time.

▲ TOP = RAW (cam[i] + tactile[i]) — tactile lags ~15f ▼ BOTTOM = CORRECTED (cam[i] + tactile[i+15])

▸ What to look for

  1. Find a contact moment in the camera row (row 1 of each panel).
  2. TOP (raw): the GelSight diff (row 2) lights up ~15 frames AFTER the camera contact.
  3. BOTTOM (corrected): the diff should light up on the SAME frame as the camera contact. If it does, +15 is right; if it's still early/late, the true shift differs.
  4. Step frames with ←/→, Space play/pause. Use 0.25× for precision.