I see that A1 predicts the frame at 15 frames, which is at 15/30=0.5s, but the chunk_size and action_size are 50, which is at 50/30 = 1.6s for RoboTwin. I'm curious why predict such a long action instead of predicting the same length action as the future video.
I see that A1 predicts the frame at 15 frames, which is at 15/30=0.5s, but the chunk_size and action_size are 50, which is at 50/30 = 1.6s for RoboTwin. I'm curious why predict such a long action instead of predicting the same length action as the future video.