---
license: mit
tags:
  - robotics
  - pi0.5
  - robocasa
---

# RACE_main_robocasa

pi0.5 checkpoints on RoboCasa H50 (25 single-stage atomic tasks, human teleop, ~55 demos/task,
1382 episodes). PyTorch (openpi `PI0Pytorch`) safetensors; norm stats included under
`assets/robocasa/robocasa_h50/`.

| folder | what | notes |
|---|---|---|
| `teacher_ac10_30k_pytorch/` | pi0.5 teacher, action_horizon 10, 30k steps from `pi05_base`, batch 64 | PyTorch export (fp32) of the JAX checkpoint in `SeonghoonYu/RACE_Robocasa/teacher_ac10_30k` |
| `ft_vlmfreeze_h10_20k/` | plain fine-tune of the teacher, VLM (SigLIP + Gemma) frozen, action expert trained, horizon 10, 20k steps | model-only (no optimizer) |
| `ft_vlmfreeze_h15_20k/` | same, horizon 15 | model-only |
| `ft_vlmfreeze_h20_20k/` | same, horizon 20 | model-only |

Fine-tune hyperparameters (all three): batch 64, constant LR 5e-5 (no warmup), AdamW, grad clip 1.0,
bf16, 20k steps, initialised from the teacher.

RoboCasa eval (25 tasks x 50 trials, held-out object split B, open-loop chunk = horizon):

| checkpoint | replan | success |
|---|---|---|
| teacher_ac10_30k | 5 | 60.4% |
| teacher_ac10_30k | 10 | 62.6% |
| ft_vlmfreeze_h10_20k | 10 | 62.5% |
| ft_vlmfreeze_h15_20k | 15 | 62.1% |
| ft_vlmfreeze_h20_20k | 20 | 58.6% |