Dataset card

README.md

@@ -13,22 +13,34 @@ configs:
   data_files: data/*/*.parquet
 ---
 
-This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
+# LatchBench SO-101 RJ45 insertion, v0
 
+Simulated demonstrations of an SO-101 arm inserting an RJ45 plug into a 3-port network switch, recorded in
+[LatchBench](https://github.com/UD-Y/latchbench) (Isaac Lab-Arena, PhysX) and stored in LeRobot v3 format with the real SO-101 conventions
+(joint names, degrees, 30 fps).
 
 <a class="flex" href="https://huggingface.co/spaces/lerobot/visualize_dataset?path=UD77/latchbench-so101-v0">
 <img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl.svg"/>
 <img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl-dark.svg"/>
 </a>
 
+## Contents
 
-## Dataset Description
+- 600 episodes, all successful (the latch clicked): 100 per port on the loose (+0.50 mm) and medium (+0.25 mm)
+  clearance tiers.
+- Recorded from the open-loop expert, given the exact port pose. Every episode is a straight push into the port. There is no search and no reaction to contact.
+- `observation.state` holds the 6 joint positions in degrees. `action` holds the 6 joint goal positions in degrees.
+- `observation.images.front` and `observation.images.wrist`: 480x640 video.
+- `observation.servo_load` and `observation.servo_current`: the simulated STS3215 `Present_Load` (signed, 0.1 % PWM
+  duty per count) and `Present_Current` (6.5 mA per count) registers, decoded the way LeRobot reads them on a real
+  arm ([servosense](https://github.com/UD-Y/servosense)).
+- `task`: "insert the cable into the left / middle / right port".
+- `meta/latchbench_episodes.jsonl`: tier, port, seed and raw-file index per episode.
 
+This is the dataset behind the v0 GR00T fine-tune, which never latched (0 of 150 evaluation episodes). The demonstrations work only because the expert knew the port position exactly, so they give a policy nothing contact-driven to learn. See the GR00T section of the LatchBench README for the full story. Model: [UD77/latchbench-groot-n17](https://huggingface.co/UD77/latchbench-groot-n17).
 
-
-- **Homepage:** [More Information Needed]
-- **Paper:** [More Information Needed]
-- **License:** apache-2.0
+Limits: everything is simulated, the plug is rigidly attached to the gripper, and mapping these joint values to a
+real arm needs that arm's LeRobot calibration file.
 
 ## Dataset Structure
 
@@ -203,11 +215,6 @@ This dataset was created using [LeRobot](https://github.com/huggingface/lerobot)
 }
 ```
 
-
 ## Citation
 
-**BibTeX:**
-
-```bibtex
-[More Information Needed]
-```
+If you use this dataset, please cite the LatchBench repository: https://github.com/UD-Y/latchbench