Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- robotics
|
| 5 |
+
- vla
|
| 6 |
+
- pi0.5
|
| 7 |
+
- openpi
|
| 8 |
+
- real-robot
|
| 9 |
+
- franka
|
| 10 |
+
- mode-editing
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
# pi05_real_pb_mixed
|
| 14 |
+
|
| 15 |
+
Fine-tuned [pi0.5](https://github.com/Physical-Intelligence/openpi) VLA model for real robot manipulation.
|
| 16 |
+
|
| 17 |
+
## Task
|
| 18 |
+
- **Task:** Push Block
|
| 19 |
+
- **Training data:** Mixed modes (from-left + from-right)
|
| 20 |
+
- **Dataset:** `real_push_block_mixed`
|
| 21 |
+
- **Robot:** Franka Panda (7-DOF)
|
| 22 |
+
- **Cameras:** Base RGB + Wrist RGB (256x256)
|
| 23 |
+
|
| 24 |
+
## Training Configuration
|
| 25 |
+
| Parameter | Value |
|
| 26 |
+
|-----------|-------|
|
| 27 |
+
| Base model | pi0.5 (PaliGemma 2B + Gemma 2B action expert) |
|
| 28 |
+
| Total parameters | ~3.35B |
|
| 29 |
+
| Action dimension | 32 |
|
| 30 |
+
| Action horizon | 10 |
|
| 31 |
+
| Batch size | 16 |
|
| 32 |
+
| Training steps | 5,000 |
|
| 33 |
+
| Learning rate | Cosine decay: warmup=500, peak=5e-5, end=5e-6 |
|
| 34 |
+
| Optimizer | AdamW (gradient clip norm=1.0) |
|
| 35 |
+
| GPUs | 8x NVIDIA A100 |
|
| 36 |
+
| Normalization | Quantile normalization |
|
| 37 |
+
|
| 38 |
+
## Checkpoints
|
| 39 |
+
- **Step 3000**: loss = 0.0071
|
| 40 |
+
- **Step 4000**: loss = 0.0047
|
| 41 |
+
- **Step 4999**
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
## Loss Curve
|
| 45 |
+
|
| 46 |
+
| Step | Loss |
|
| 47 |
+
|------|------|
|
| 48 |
+
| 0 | 0.0818 |
|
| 49 |
+
| 500 | 0.0155 |
|
| 50 |
+
| 1000 | 0.0120 |
|
| 51 |
+
| 1500 | 0.0110 |
|
| 52 |
+
| 2000 | 0.0084 |
|
| 53 |
+
| 2500 | 0.0075 |
|
| 54 |
+
| 3000 | 0.0071 |
|
| 55 |
+
| 3500 | 0.0060 |
|
| 56 |
+
| 4000 | 0.0047 |
|
| 57 |
+
| 4500 | 0.0042 |
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
## Part of Mode Editing Research
|
| 61 |
+
This checkpoint is part of the "Don't Filter Your Data, Edit Your Policy" project (CoRL 2026),
|
| 62 |
+
investigating post-hoc behavior mode editing for robot policies using Classifier-Guided Distillation (CG-Distill).
|