README.md

---
license: apache-2.0
---

# Mem-0 Execution Module — Mn tasks (RMBench / RoboTwin 2.0)

**Per-task Mem-0 execution-module checkpoints** for four RMBench **Mn** tasks:
`battery_try`, `blocks_ranking_try`, `cover_blocks`, `press_button`.
Each task has its **own** executor — trained only on that task's 50-episode
`demo_clean` dataset — and its **own** normalization stats (unlike the
[`m1_mix` release](https://huggingface.co/qiuly/Mem-0-m1mix-RMBench), which used one
joint checkpoint for all five M1 tasks).

- **Backbone:** Qwen3-VL-2B-Instruct (fine-tuned; weights bundled in each checkpoint)
- **Action head:** DiT-B flow-matching policy (16-D action)
- **Memory:** MemoryBank (instant + anchor memory fusion across the episode)
- **Aux head:** subtask-end classifier
- ≈ 2.67 B params per checkpoint; 30,000 training steps per task

## Results

`task_config = demo_clean`, `instruction_type = unseen`, 100 episodes per task.
`action_horizon` is the per-task inference-time setting used for these numbers.

| 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**    | —      |

Rollout videos (100 per task) and raw score files are under `eval_results/`
(see `eval_results/results.md`).

## Contents

```
mn_submit/
├── README.md
├── checkpoints/
│   └── <task>/final_step30000.pt (+ .sha256)     # one full ckpt per task, ~15.3 GB each
├── norm_stats/
│   └── <task>/norm_stats.json                    # per-task min-max state/action stats
└── eval_results/
    ├── results.md                                # scores + per-task action_horizon
    └── <task>/                                   # _result.txt, episode*.mp4 (×100)
```

## Checkpoints

Each `checkpoints/<task>/final_step30000.pt` is a **single, unsplit** full training
checkpoint (~15.3 GB): `model_state_dict` + `optimizer_state_dict` +
`scheduler_state_dict`, `global_step` 30000. Verify integrity with the `.sha256`
file next to it:

```bash
sha256sum -c final_step30000.pt.sha256
```

The `model_state_dict` is self-contained — it already includes the fine-tuned
Qwen3-VL-2B backbone weights. Optional inference-only slimming (~15.3 GB → ~6 GB):

```python
import torch
ck = torch.load("final_step30000.pt", map_location="cpu", weights_only=False)
torch.save({"model_state_dict": ck["model_state_dict"], "global_step": ck["global_step"]},
           "final_step30000_inference.pt")
```

## Base VLM

Model instantiation requires the official `Qwen/Qwen3-VL-2B-Instruct` directory
(architecture + tokenizer/processor; Apache-2.0). Its weights are overwritten by the
checkpoint at load time, but the directory must exist locally:

```bash
huggingface-cli download Qwen/Qwen3-VL-2B-Instruct \
    --local-dir policy/Mem-0/checkpoints/Qwen3-VL-2B-Instruct
```

## Normalization

State and action are min-max normalized to `[-1, 1]` using the per-task stats in
`norm_stats/<task>/norm_stats.json`. Always pair a checkpoint with **its own task's**
stats at inference; predicted actions are denormalized with the same file.

## License & attribution

Base VLM **Qwen3-VL-2B-Instruct** is © the Qwen team, licensed **Apache-2.0**; the
checkpoints embed fine-tuned Qwen weights, so that license applies to the
corresponding components. RMBench / RoboTwin and the Mem-0 policy code are governed
by their respective upstream licenses.