TLDR Gripe / News
AIRoA-0 Learns Complex Robot Tasks from Human Handheld Videos
A new system called AIRoA-0 has been developed by researchers to master intricate, extended manipulation tasks for robots. It was trained exclusively on 16,151 hours of video showing humans performing tasks with handheld devices, without using any robot-specific data. Using only wrist-worn cameras, the system can execute sophisticated procedures on different robot bodies, achieving a 78% success rate in ten different evaluation scenarios. The development team is making the training data, model weights, and instructions for further training available to the public.
Read the full story