Language-Table

Language-Table is a suite of demonstration datasets and a continuous-control benchmark for open-vocabulary visuolinguomotor learning, released alongside the paper 'Interactive Language: Talking to Robots in Real Time' (Lynch et al.). The flagship real-robot split contains 442,226 episodes of teleoperated tabletop manipulation with hindsight-relabeled natural-language instructions such as 'push the blue triangle to the top left corner'; across all nine released variants (real + simulation, human + oracle) there are 1,639,544 episodes. Each episode is a sequence of steps whose observation is a 360x640 RGB image plus the 2D end-effector translation and a UTF-8-encoded instruction, and whose action is a 2D delta Cartesian setpoint for the arm. Data is distributed openly under Apache-2.0 from the public gs://gresearch/robotics/ Google Cloud Storage bucket in RLDS (TFDS/TFRecord) format. The demonstrations are human teleoperation with no dense per-step reward; the benchmark defines success per task family (e.g. a pushed block reaching a distance threshold of its target), and the paper reports a 93.5% human-judged success rate over 87,000 unique instructions.