---
tags: [robotics, vla, pi0.5, manipulation, hsr, icra2026]
license: apache-2.0
---
# Pi0.5 HSR V1
- Base: Pi0.5 (~2.3B, JAX/Flax)
- Training: 50K steps, bs=32, 2xH100, ~21h
- Dataset: 70K episodes curated from 2.5M
- Checkpoint: step 49999 (Orbax)
- State: joints(8) matches eval