Chinese robotics startup LiberAI has released Liber-0, a World Action Model designed to control robot limbs and hands without the need for task-specific training. In testing across 60 unfamiliar tasks and 600 trials, the model achieved a 36% overall success rate, with performance reaching 63.5% for pick-and-place operations and 10.5% for complex multi-step sequences. A similar test of GPT-6 Astra resulted in a 5% completion rate for the same set of tasks.
The model was pretrained on more than 2.2M hours of human experience, leveraging motion data from the company’s 131 g LiberSense sensor glove. LiberAI reported that 10,000 hours of sensor glove data outperformed 2M hours of web video in matched runs. The startup says using human sensor data instead of having humans manually steer robots through tasks could lower the cost and time required for training.
