Goal: make a meaningful manipulation task that can be trained in under 10 mins on a large GPU

Lerobot pushT:

• Image-based: 91% SR in 6 hours with ~5.4 GB VRAM https://www.notion.so/robot-learning/PushT-reproduce-diffusion-28c0055427458057a206ef9c58e695e4
• Vanilla keypoint-based: % SR in hours with ~5.2 GB VRAM https://wandb.ai/pokeandwiggle/pusht_flash/runs/u6bqr328?nw=nwuserzaringleb

Original keypoints pushT:

batch size: 256, n_obs: 2, 0.916% SR in ~17 mins, 1.7 GB VRAM