---
license: apache-2.0
tags:
- robotics
- lerobot
- molmoact2
- so101
- diagnostic
---

# MolmoAct2 Banana Diagnostic - frombase prod r128 c010 rw3 w8

Diagnostic MolmoAct2 checkpoint for SO-101 banana color-conditioned pick-and-place.

**Important:** this checkpoint is for directionality/robot diagnostic testing only. It did **not** pass the strict offline robot-ready gates from the overnight evaluation.

## Source

- Run: `molmoact2_overnight_frombase_prod_r128_c010_rw3_w8_gpu5_20260519_fullcoverage_v1`
- Checkpoint step: `010000`
- Dataset: `rslxcvg/banana_act_direct_color_simple_v1_molmo_compat`
- Base checkpoint: `allenai/MolmoAct2-SO100_101`
- Norm tag: `so100_so101_molmoact2`
- Gripper normalization: enabled

## Offline Metrics

| metric | value |
|---|---:|
| mixed rank | 1 |
| low_t accuracy | 0.889 |
| pan_rank | 1.000 |
| pan5 degrees | 18.937 |
| final_pan degrees | 61.162 |
| outside stats pct | 0.000 |
| horizon gripper MAE | 1.205 |

Strict gates were `low_t >= 0.95`, `pan_rank == 1.0`, `pan5 >= 20`, `outside == 0`, and `h_grip <= 2.5`. This checkpoint failed the low_t and pan5 gates.

## Files

The repository root is the LeRobot/MolmoAct2 `pretrained_model` directory and should be usable as a policy checkpoint path after `snapshot_download`.

Evaluation artifacts are under `eval/`.