Language-Table
Robotics at Google / Google Research · 2022 · datasets.bot · datasets.bot page
One-liner. Google's large language-conditioned tabletop block-manipulation dataset: ~442K real + ~181K sim xArm6 trajectories, plus a sim benchmark.
Setup
- Datasets / benchmarks: Language-Table is a suite of human-collected datasets and a multi-task continuous control benchmark for open-vocabulary visuolinguomotor learning, released with the paper 'Interactive Language: Talking to Robots in Real Time'. It contains nearly 600,000 natural-language-labeled trajectories of a UFACTORY xArm6 robot moving colored blocks on a tabletop (442,226 real-robot episodes plus simulated human-controlled and oracle episodes). Observations are third-person RGB images (captured at 640x360, resized to 320x180) plus end-effector/block state, and the arm is constrained to a 2D plane and acts via a 2D Cartesian end-effector delta setpoint. License: Apache-2.0. Download: https://github.com/google-research/language-table.
- Hardware / simulator: Embodiment: xarm. Environment: simulation, tabletop. Realness: both.
Schema
RLDS/TFDS: episodes -> steps -> {observation: {rgb 320x180, effector_translation, effector_target_translation}, action: 2D EE delta, instruction (language)}
Links