China’s LiberAI Reports 36% Success Rate for New Robot Model

BEIJING - Robotics company LiberAI has launched its Liber-0 AI, an AI model that has been trained with more than 2.2 million hours of human activity data and has shown a success rate of 36% in its trials on tasks that have not been taught explicitly to the robots.

The company reported that the trial was carried out 600 times on 60 manipulation tasks where each of the tasks was tried 10 times. The success rate for picking up and placing object task was 63.5%, whereas it dropped to 34% for other shorter manipulation tasks and 10.5% for complicated multistep manipulation tasks.

Other robot foundation models also fail in executing unfamiliar tasks. Nvidia showed a success rate of 38.3% with its GR00T N1.5 model when trying to carry out 12 simulation manipulation tasks; however, the model was pre-trained with synthetic trajectory data on those manipulation tasks. The two results cannot be compared as LiberAI has tested with physical robots.

Liber-0 uses videos of human activities, motion data through LiberAI’s LiberSense Glove, and robot demonstration data. Unlike typical video, glove recording gives measurements of how humans manipulate their hands when interacting with objects, which helps the AI to correlate actions with movements a robot can make.

LiberAI performed tests of the model on its LEO robot with dexterous hands, articulated wrists, and cameras set up to mimic human observation angles. Company's separate tests show that the combination of glove recordings and fewer robot demonstrations yields higher manipulation scores than the robot demonstrations alone. LiberAI claims to have used 75% less robot demonstrations in such test.

Commercial applications of Liber-0 have not been disclosed yet. All the statistics given by the company Nov. 3 have not been independently verified.