Add half-subset baseline 12-task eval (pi05half, pi0half; xvla pending)
Browse files- half_models/README.md +11 -2
half_models/README.md
CHANGED
|
@@ -1,3 +1,12 @@
|
|
| 1 |
-
# Half-subset baselines
|
| 2 |
|
| 3 |
-
pi0.5
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Half-subset baselines — 12-task RoboPRO eval
|
| 2 |
|
| 3 |
+
Models trained on the roboreal 50-clean+50-clutter subset (pi0.5 20k, pi0 40k), evaluated on the 12 curated tasks: clean (20 seeds/task) + clutter (10 densities x2 = 20/task). Eval config matches training: adapt_to_pi=False, use_delta_joint_actions=True.
|
| 4 |
+
|
| 5 |
+
| Model | Clean SR | Clutter SR |
|
| 6 |
+
|---|---|---|
|
| 7 |
+
| pi0.5-half-20k | 78% | 50% |
|
| 8 |
+
| pi0-half-40k | 12% | 12% |
|
| 9 |
+
|
| 10 |
+
**xvla-half-60k excluded**: its eval returned 0% across all 480 episodes — a load failure (EE6D pipeline), NOT a real result. Under investigation; will be added once fixed.
|
| 11 |
+
|
| 12 |
+
Note: pi0-half SR (12%) is low; likely an under-trained subset model, pending a sanity re-check.
|