# PI0.5 few-shot checkpoint: `global_step_14000`

This directory contains the RLinf-native PI0.5 checkpoint saved at training
step **14,000** from the `Pi05_few_shot_libero` few-shot run. It is an
independent, inference-ready checkpoint and includes its own normalization
statistics and consolidated LIBERO evaluation results.

## Model description

- Base policy: PI0.5 / OpenPI, config name `pi05_libero`
- Training data: 10 selected episodes per task, sampling seed `42`
- Suites represented in the run: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal,
  and LIBERO-10
- Action dimension: `7` (six robot motion dimensions plus one gripper
  dimension)
- Gripper convention: normalized value **`0 = close`, `1 = open`**
- Inference: 10 denoising steps, noise level `0.5`, five action chunks per
  inference

The gripper convention is part of the model/evaluation contract. Keep the
seventh action value in `[0, 1]` and do not reverse its meaning in a custom
adapter.

## Evaluation result

The checkpoint was evaluated on all four suites with 50 episodes per task,
500 trajectories per suite, fixed ordered reset states, and evaluation seed
`195`. The aggregate success-once rate is **0.713**
(1425/2000).

| Suite | Trajectories | Success-once | Success-at-end |
|---|---:|---:|---:|
| LIBERO-Spatial | 500 | 0.746 | 0.160 |
| LIBERO-Object | 500 | 0.862 | 0.120 |
| LIBERO-Goal | 500 | 0.774 | 0.156 |
| LIBERO-10 | 500 | 0.468 | 0.110 |

The complete reusable parameters are in [`eval/summary.json`](./eval/summary.json),
and per-task results are in the suite subdirectories under `eval/`.

## Files and loading

- `actor/model_state_dict/full_weights.pt`: complete model state dict
- `physical-intelligence/libero/norm_stats.json`: required normalization
  statistics for this checkpoint
- `eval/summary.json`: evaluation parameters and aggregate metrics
- `rlinf_model_config.yaml`: portable RLinf model fragment

Set the checkpoint directory before evaluation:

```bash
export PI05_FEWSHOT_CHECKPOINT_DIR=/absolute/path/to/Pi05_few_shot_libero/global_step_14000
```

Then use the included [`rlinf_model_config.yaml`](./rlinf_model_config.yaml) and
keep `openpi_data.norm_stats_path` pointed at the normalization file in this
directory.