Language-Table

Language-Table is a large-scale dataset of language-annotated robot manipulation trajectories collected by Google Research. Built on a Franka Panda robot arm operating on a tabletop, the dataset contains over 442,000 demonstration trajectories, each paired with natural language instructions describing the manipulation task. Trajectories cover diverse tabletop interactions including pushing, picking, placing, and rearranging colored blocks. Language annotations range from simple commands ('push the red block to the left') to more complex relational instructions ('move the green cube behind the blue cylinder'). The dataset enabled the development of interactive language-conditioned policies that can understand and execute natural language commands in real-time robotic settings. It has been used to train models capable of zero-shot generalization to new objects and instructions, and remains a standard benchmark for language-conditioned robotic manipulation.