SSSSphinx commited on
Commit
0b85c5b
·
1 Parent(s): 24f777e

Publish only the LIBERO updater

Browse files
Files changed (1) hide show
  1. README.md +7 -10
README.md CHANGED
@@ -5,17 +5,14 @@ tags:
5
  - safetensors
6
  ---
7
 
8
- # CauSyncVLA updater weights
9
 
10
- These are the two trained KV updaters used in *CauSyncVLA: Causal Latent World Modeling for Asynchronous Robot Control*. The [code and loading instructions](https://github.com/TwSphinx54/CauSyncVLA) include the frozen-policy adapters, evaluation entry points and training recipes.
11
 
12
- | File | Matching frozen expert | Training step | History | Feedback |
13
- | --- | --- | ---: | ---: | --- |
14
- | `libero/updater.safetensors` | OpenPI π0.5-LIBERO | 12,000 | 48 | Executed simulator steps |
15
- | `franka/updater.safetensors` | Laboratory π0.5 FR3 expert | 8,500 | 24 | Time-conditioned physical feedback |
16
 
17
- Each file includes the learned tensors and original KV/state normalization buffers. The architecture, global training step, and expected action-normalization identity are stored in safetensors metadata. Both exports were read back and checked against every tensor of the corresponding training checkpoint.
18
 
19
- The LIBERO expert is available through [OpenPI](https://github.com/Physical-Intelligence/openpi). The FR3 expert and its action normalization are privately held. Each adapter requires its matching frozen model and input convention. Reproduction settings and outcome denominators are recorded in the code repository.
20
-
21
- The original FR3 trajectories were filtered at collection cadence. Timestamped augmentation for updater training preserves the original issued actions, marks synthetic intermediate states and separates gripper events. The physical profile uses its corresponding expert's original action scale.
 
5
  - safetensors
6
  ---
7
 
8
+ # CauSyncVLA LIBERO updater
9
 
10
+ The LIBERO KV-cache updater from *CauSyncVLA: Causal Latent World Modeling for Asynchronous Robot Control* conditions a frozen OpenPI π0.5-LIBERO expert on the latest published visual cache and executed state/action history.
11
 
12
+ | File | Training step | History |
13
+ | --- | ---: | ---: |
14
+ | `libero/updater.safetensors` | 12,000 | 48 simulator steps |
 
15
 
16
+ The safetensors file contains the learned parameters and original KV/state normalization buffers. Its metadata records the architecture and expected action-normalization identity. The export was checked against every tensor of the training checkpoint.
17
 
18
+ Download the matching frozen expert through [OpenPI](https://github.com/Physical-Intelligence/openpi). [CauSyncVLA code and evaluation instructions](https://github.com/TwSphinx54/CauSyncVLA) describe installation, checkpoint conversion and reproduction of the LIBERO evaluation.