minliu01 commited on
Commit
5459c73
·
verified ·
1 Parent(s): 614eaa4

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +7 -8
README.md CHANGED
@@ -8,14 +8,13 @@ size_categories:
8
  - 100M<n<1B
9
  ---
10
  # ManipGen-PartNet
11
- ManipGen-PartNet is used to train pick and place policies in [ManipGen](https://mihdalal.github.io/manipgen/). We process the [UniDexGrasp](https://pku-epic.github.io/UniDexGrasp/) dataset with the following steps:
12
- 1. Augment object size: We add smaller versions of the objects to the dataset.
13
- 2. Augment rest poses: The rest object poses in the original dataset are obtained by dropping the object from a height. We add some common rest poses which are less likely to be sampled in the original dataset (e.g., a standing bottle).
14
- 3. Sample grasp pose: We sample grasp poses for each object using antipodal sampling. Then we teleport the gripper to grasp the object from the sampled poses, add perturbations, and record valid grasp poses. The ratio of valid grasp pose is also used to filter intractable rest poses (e.g., an upside-down bowl).
15
 
16
  ### Structure
17
- * `datasetv4.1_posedata.npy`: rest poses
18
- * `meshdatav3_scaled/`: mesh and urdf
19
- * `graspdata/`: pre-sampled grasp poses
20
- * `trainset3419.txt`: the list of 3419 objects used to train pick and place policies in ManipGen in the format of "[object_code] [object_scale]"
21
 
 
8
  - 100M<n<1B
9
  ---
10
  # ManipGen-PartNet
11
+ ManipGen-PartNet is used to train grasp handle, open, and close policies in [ManipGen](https://mihdalal.github.io/manipgen/).
12
+
13
+ The dataset contains 2K+ cabinet assets (drawers and doors) for training robotic manipulation policies in simulation. We sample 1K+ handles from [PartNet](https://partnet.cs.stanford.edu/) and assemble them with procedually generated cabinet bodies.
14
+
15
 
16
  ### Structure
17
+ * `meshdata`: mesh and urdf
18
+ * `graspdata/`: pre-sampled grasp poses for Franka arm with UMI gripper
19
+ * `trainset3419.txt`: the list of 2655 objects used to train grasp handle, open, and close policies in ManipGen
 
20