Add README
Browse files
README.md
ADDED
|
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- robotics
|
| 5 |
+
- manipulation
|
| 6 |
+
- pi0
|
| 7 |
+
- hsr
|
| 8 |
+
- icra2026
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# Pi0.5 HSR V2 — ICRA 2026 Mobile Manipulation
|
| 12 |
+
|
| 13 |
+
Fine-tuned Pi0.5 (~2.3B params) for the Toyota HSR (Human Support Robot) on 80K GIST-curated episodes.
|
| 14 |
+
|
| 15 |
+
## Training Details
|
| 16 |
+
|
| 17 |
+
| Param | Value |
|
| 18 |
+
|-------|-------|
|
| 19 |
+
| Base model | Pi0.5 (gs://openpi-assets/checkpoints/pi05_base) |
|
| 20 |
+
| Framework | JAX/Flax via openpi |
|
| 21 |
+
| Dataset | 80,719 episodes, 14.2M frames, 93 task types, 49 robots |
|
| 22 |
+
| Curation | GIST (FAISS max-min diversification, 2-level allocation) |
|
| 23 |
+
| Steps | 50,000 |
|
| 24 |
+
| Batch size | 32 |
|
| 25 |
+
| LR | 5e-5 (cosine decay, 2K warmup) |
|
| 26 |
+
| EMA | 0.99 |
|
| 27 |
+
| ProprioDropout | 0.4 |
|
| 28 |
+
| Action horizon | 16 |
|
| 29 |
+
| Hardware | 2× H100 PCIe 80GB |
|
| 30 |
+
| Training time | ~23h |
|
| 31 |
+
| Final loss | 0.044 |
|
| 32 |
+
|
| 33 |
+
## Checkpoint Format
|
| 34 |
+
|
| 35 |
+
OpenPI / Orbax checkpoint format (ocdbt). Load with:
|
| 36 |
+
|
| 37 |
+
```python
|
| 38 |
+
from openpi.training import config as _config
|
| 39 |
+
config = _config.get_config("pi05_hsr_v3")
|
| 40 |
+
```
|
| 41 |
+
|
| 42 |
+
## Files
|
| 43 |
+
|
| 44 |
+
- `params/` — Model parameters (EMA-averaged)
|
| 45 |
+
- `train_state/` — Full optimizer state (for resuming)
|
| 46 |
+
- `assets/` — Norm stats for the dataset
|