PI0.5 LIBERO-10 task-8 initial policy
This checkpoint is the deliberately weak initial policy used in the ReCap
study on LIBERO-10 task 8, “put both moka pots on the stove.” It starts from
Miical/pi05-base and is fully
fine-tuned on 10 official task demonstrations for 20 epochs (300 optimizer
steps).
Training configuration
| Setting | Value |
|---|---|
| Training episodes | 10 |
| Training frames | 4,033 |
| Epochs / optimizer steps | 20 / 300 |
| Model GPUs | 8 |
| Global / micro batch size | 256 / 16 |
| Learning rate | 1e-4 |
| Weight decay | 1e-5 |
| Warmup | 5% |
On the complete 50-episode task-8 evaluation, this policy succeeds on 8/50 episodes (16%). The intentionally moderate success rate leaves room to measure post-training improvement. Three ReCap iterations raise the best measured success rate to 23/50 (46%).
The associated cumulative dataset is
Miical/pi05-libero10-task8-recap.
Its first 10 episodes are the demonstrations used for this checkpoint.