Document RoboTwin action statistics
Browse files
README.md
CHANGED
|
@@ -50,6 +50,12 @@ Each native bundle contains three matching components:
|
|
| 50 |
and `prior.pt` contain the subgoal generator and GMM action prior separately.
|
| 51 |
Keep all components from the same suite together.
|
| 52 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
## Usage
|
| 54 |
|
| 55 |
Download the checkpoints with the Hugging Face Hub:
|
|
|
|
| 50 |
and `prior.pt` contain the subgoal generator and GMM action prior separately.
|
| 51 |
Keep all components from the same suite together.
|
| 52 |
|
| 53 |
+
Action normalization statistics accompany the native bundles: `action_stats.json`
|
| 54 |
+
for LIBERO and [`robotwin_a2b/action_stats.npz`](robotwin_a2b/action_stats.npz)
|
| 55 |
+
for RoboTwin A2B. The RoboTwin file contains the paper's 14-dimensional `mean`
|
| 56 |
+
and `std` arrays and `count=1188823`. Use the released statistics without
|
| 57 |
+
recomputing them; the GitHub model installer includes and verifies this file.
|
| 58 |
+
|
| 59 |
## Usage
|
| 60 |
|
| 61 |
Download the checkpoints with the Hugging Face Hub:
|