SKU DEM-6004 · Sold by External
CALVIN
Product specifications
| SKU | DEM-6004 |
|---|---|
| Data type | Demonstrations |
| Volume | 24 hours of teleoperated play data (6 h in each of 4 environments A/B/C/D); 34 annotated tasks |
| Size on disk | Split-dependent: D->D 166 GB, ABC->D 517 GB, ABCD->D 656 GB; debug 1.3 GB |
| Format | Per-timestep NumPy .npz frames + .npy language-annotation files (episode_*.npz, lang_annotations/auto_lang_ann.npy); PyBullet simulation |
| Access model | PUBLIC LICENSE |
| Pricing | Free · open-source license (MIT) |
| Quality score | — |
| License | MIT License |
CALVIN (Composing Actions from Language and Vision) is a PyBullet-simulated benchmark for language-conditioned policy learning for long-horizon robot manipulation, released with the mees/calvin repository under the MIT license. The dataset consists of unstructured, human-teleoperated 'play' data: 6 hours in each of four visually distinct table-top environments (A, B, C, D), totalling ~24 hours, with a Franka Emika Panda 7-DoF arm. Each interaction timestep is stored in a NumPy file containing static and gripper RGB images, depth maps, a tactile image, 7-DoF absolute and relative actions, 15-D robot proprioceptive state, and 24-D scene state. Short windows of the play data are annotated via crowd-sourcing with free-form language instructions covering 34 tasks (rotating, pushing, lifting, stacking, sliding drawers/doors, and toggling lights). The benchmark's Long-Horizon Multi-Task Language Control (LH-MTLC) protocol chains sequences of five instructions, and success is judged by a rule-based oracle task detector in calvin_env rather than by any reward label in the data itself. It is distributed as three train/eval splits (D->D, ABC->D, ABCD->D) plus a small debug set.