SteveNguyen commited on
Commit
05e70fd
·
verified ·
1 Parent(s): 8e7f192

Add grabette tags and the projection/chunk-relative note

Browse files
Files changed (1) hide show
  1. README.md +28 -1
README.md CHANGED
@@ -4,6 +4,10 @@ task_categories:
4
  - robotics
5
  tags:
6
  - LeRobot
 
 
 
 
7
  configs:
8
  - config_name: default
9
  data_files: data/*/*.parquet
@@ -150,4 +154,27 @@ This dataset was created using [LeRobot](https://github.com/huggingface/lerobot)
150
 
151
  ```bibtex
152
  [More Information Needed]
153
- ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4
  - robotics
5
  tags:
6
  - LeRobot
7
+ - grabette
8
+ - gripette
9
+ - grasp-projection
10
+ - chunk-relative
11
  configs:
12
  - config_name: default
13
  data_files: data/*/*.parquet
 
154
 
155
  ```bibtex
156
  [More Information Needed]
157
+ ```
158
+
159
+ ## Grabette notes
160
+
161
+ Gripper channels are **(strategy, closure)**, not joint angles: the last two
162
+ channels of `action` and `observation.state` are normalised to the gripper's
163
+ reachable travel. In `action`, `closure = 1.0` means "drive fully closed" and
164
+ the object stops the fingers; in `observation.state` closure stays continuous,
165
+ so "am I holding something" remains observable. The exact calibration is in
166
+ `meta/grasp_projection.json`.
167
+
168
+ `action` is **8D absolute pose** `[x, y, z, ax, ay, az, strategy, closure]` —
169
+ deliberately *not* per-step deltas. `meta/stats.json` carries
170
+ `action_relative_meta`, so the action stats describe the **chunk-relative**
171
+ distribution (chunk_size 50) rather than the absolute one. Training must set
172
+ `GRABETTE_CHUNK_RELATIVE=1`; `integrations/Pi05/train.py` refuses otherwise
173
+ rather than normalise with the wrong scale.
174
+
175
+ Prepared from [chouziel/grabette-sugar-cup-2008](https://huggingface.co/datasets/chouziel/grabette-sugar-cup-2008)
176
+ to be channel-for-channel comparable with
177
+ [chouziel/grabette-sugar-cup-2008_graspproj](https://huggingface.co/datasets/chouziel/grabette-sugar-cup-2008_graspproj)
178
+ (the per-step-delta baseline): the projected gripper channels are bit-identical
179
+ across all 23046 frames, so an A/B between them isolates the action
180
+ representation.