One Unitree model handles 64 tasks, but its reliability is still in question.

Unitree G1, controlled by a single model, performs 64 tasks, from working at a desk to full-body actions. It sounds like a step toward more unified humanoid control, but the list of demonstrations does not answer the main question: how reliably does the robot repeat actions when the environment changes?

The model itself, UnifoLM-WLA-1.0, contains 6 billion parameters and was trained on approximately 2,500 hours of data from real robots. It supports parallel grippers and two types of dexterous hands. The demonstrated tasks include making a bed, loading a washing machine, and folding clothes.

What is interesting about the concept: the model combines spatial reasoning with motion control. First, it assesses the situation, then predicts changes in the scene and generates actions for the arms and lower body. This could potentially be more useful than a set of separate narrow models, making it easier for the robot to coordinate actions that require moving and manipulating objects simultaneously.

Still, 64 tasks do not mean 64 reliable household skills. The available description contains no specific success rates or details of independent verification. The claim of strong test results is also presented without figures. For now, this is a convincing step toward universal control, but not proof that such a robot can be sent off to do household chores unsupervised.

What task would you trust a humanoid robot with first if it could complete it only most of the time?

Source: LocalLlama