Latent action for rotation

ALOHA demonstrations

Ablation study on model architecture

(1) using Conv2D to extract latent action

(2) using readout token to extract latent action

(3) using single ViT to extract latent action

Compare with the UniPi baseline