Former DeepMind research scientist releases an embodied model that can adapt to new tasks in 13 minutes.

On August 11, DynaRobotics, a company founded by three Chinese entrepreneurs in the United States, released its next-generation robotic foundation model Dyna-2, achieving a success rate of over 80% in high-precision manufacturing tasks. The success rate of task execution and generalization ability are currently two core issues for robots. This model, pre-trained on over 1 million hours of human video, requires only about 13 minutes of real machine remote operation demonstration data to adapt to new tasks, such as using two dexterous hands to twist bottle caps; the success rate for high-precision manufacturing tasks has increased from 20% in the previous model Dyna-1 to 80%-90%. In deployment scenarios, Dyna-2's pass rate is 87%, while the pass rate for the previous model Dyna-1 was only 46%.