Pi0.5 HSR V2 — ICRA 2026 Mobile Manipulation
Fine-tuned Pi0.5 (~2.3B params) for the Toyota HSR (Human Support Robot) on 80K GIST-curated episodes.
Training Details
| Param | Value |
|---|---|
| Base model | Pi0.5 (gs://openpi-assets/checkpoints/pi05_base) |
| Framework | JAX/Flax via openpi |
| Dataset | 80,719 episodes, 14.2M frames, 93 task types, 49 robots |
| Curation | GIST (FAISS max-min diversification, 2-level allocation) |
| Steps | 50,000 |
| Batch size | 32 |
| LR | 5e-5 (cosine decay, 2K warmup) |
| EMA | 0.99 |
| ProprioDropout | 0.4 |
| Action horizon | 16 |
| Hardware | 2× H100 PCIe 80GB |
| Training time | ~23h |
| Final loss | 0.044 |
Checkpoint Format
OpenPI / Orbax checkpoint format (ocdbt). Load with:
from openpi.training import config as _config
config = _config.get_config("pi05_hsr_v3")
Files
params/— Model parameters (EMA-averaged)train_state/— Full optimizer state (for resuming)assets/— Norm stats for the dataset