AIRoA-0 manipulation system trained from handheld data

Researchers introduce AIRoA-0, a bimanual manipulation system trained entirely on 16,151 hours of handheld demonstrations without any robot data. The system performs complex long-horizon tasks across various robot embodiments using only wrist cameras and achieves an average success rate of 78% across ten evaluation tasks. The creators are releasing the dataset, pre-trained weights, and post-training recipe publicly. Open article →