Researchers introduced PointZero, a method for 3D point track completion. It learns transferable 3D dynamics without robot action labels. They built a dataset of 2.9 million synthetic frames. PointZero uses a transformer architecture. It outperforms prior methods on the same data. Fine-tuned PointZero beats baselines on the PGND 3D dynamics benchmark. It also outperforms or matches baselines on 6 of 7 robot manipulation tasks. The dataset, checkpoints and training recipe are released. The work targets action-conditioned 3D dynamics prediction and imitation learning.

• 2.9 million synthetic frames
• 6/7 robot manipulation tasks outperformed or matched