eval_results/results.md

# Mn Evaluation Results

`task_config = demo_clean`, `instruction_type = unseen`, 100 episodes per task.
Each task is evaluated with its own separately trained execution-module checkpoint
and its own normalization stats.

| Task               | action_horizon | Success Rate | Reward |
|--------------------|:--------------:|:------------:|:------:|
| battery_try        | 15             | **0.21**     | 0.00   |
| blocks_ranking_try | 15             | **0.15**     | 0.00   |
| cover_blocks       | 8              | **0.76**     | 0.82   |
| press_button       | 30             | **0.02**     | 0.00   |
| **Average**        | —              | **0.285**    | —      |

## Action horizon

Per-task inference-time `action_horizon` settings:

- `battery_try` — **15**
- `blocks_ranking_try` — **15**
- `cover_blocks` — **8**
- `press_button` — **30**

All results above were produced with these values.

## Files

- `<task>/episode0.mp4 … episode99.mp4` — full rollout videos (100 per task)
- `<task>/_result.txt` — raw score file for that run