Can robots learn to handle fragile objects gently from language and touch?

Robots struggle to handle delicate objects safely because they lack good touch-based feedback during learning. Tabero solves this by converting existing robot videos into vision-tactile-language training data and introducing a hybrid control system that decouples force commands from position commands. On gentle manipulation tasks, the resulting model cuts grip force by over 70% while maintaining task success—showing robots can learn to be deliberately careful based on language instructions and real-time force sensing.