RSS 2026 Post-Training pi0.5 Models
OpenPI pi0.5 checkpoints used by Team UTokyo for the RSS 2026 Post-Training for Robotics Foundation Models Challenge.
Technical report: https://takehirohara.github.io/rss_2026_workshop_posttraining/
Checkpoints
Phase 1
| Path | Policy | Training step |
|---|---|---|
phase1/generalist/ |
Multi-task generalist | 190000 |
phase1/mouse-specialist/ |
insert-mouse-battery specialist |
49999 |
phase1/hanoi-specialist/ |
tower-of-hanoi-game specialist |
49999 |
phase1/seal-specialist/ |
seal-water-bottle-cap specialist with post-success reset removal |
99999 |
Phase 2
| Path | Policy | Additional post-training steps |
|---|---|---|
phase2/generalist/ |
Multi-task generalist initialized from the Phase 1 349999 checkpoint | 60000 |
phase2/hanoi-specialist/ |
tower-of-hanoi-game specialist initialized from the Phase 1 190000 checkpoint |
40000 |
phase2/seal-specialist/ |
seal-water-bottle-cap specialist initialized from the Phase 1 99999 checkpoint |
40000 |
The Phase 2 insert-mouse-battery submission reused the Phase 1 mouse specialist because its Phase 2 post-training did not finish before submission.
Format
Each directory contains an Orbax params checkpoint, normalization statistics under assets, and checkpoint metadata. Optimizer and training-resume state (train_state) is intentionally omitted.
Use these checkpoints with the OpenPI baseline referenced in the technical report. The checkpoints are derived from the pi0.5 base model and remain subject to the applicable upstream OpenPI/pi0.5 terms and licenses.