Add dataset card

README.md

@@ -0,0 +1,27 @@
+---
+license: other
+task_categories:
+- robotics
+pretty_name: Image CCIL experiment datasets
+---
+
+# Image CCIL experiment datasets
+
+Datasets used by the Image CCIL PushT and Robomimic experiments.
+
+| Path | Task | Episodes | Steps | Notes |
+|---|---|---:|---:|---|
+| `pusht/pusht_cchi_v7_replay_41.zarr/` | PushT | 41 | 5,216 | Experiment dataset |
+| `can/can_ph_demo_v141_20_perc.hdf5` | Can | 40 | 4,542 | 20% source subset |
+| `square/square_abs_ph_demo_v141_20_perc.hdf5` | Square | 40 | 6,010 | 20% source subset |
+| `tool_hang/tool_hang_all_success_ph_demo_v141_20_perc.hdf5` | ToolHang | 40 | 18,100 | 20% source subset |
+| `transport/transport_ph_demo_v141_20_perc.hdf5` | Transport | 40 | 19,045 | 20% source subset |
+
+The Robomimic 10% experiments select the first 20 demonstrations at load time
+from these 40-trajectory files (for example, with `dataset.num_traj=20` or
+`--max_episodes 20`). The files themselves contain all 40 trajectories listed
+above.
+
+PushT is stored as an unpacked Zarr directory. The Robomimic tasks are stored as
+HDF5 files and retain the observation, action, and environment metadata used by
+the training and fixed-state evaluation pipelines.