File size: 2,615 Bytes
a72c373
1bcb99b
 
 
a72c373
 
 
1bcb99b
 
a72c373
 
1bcb99b
a72c373
b02a025
 
1bcb99b
a72c373
1bcb99b
a72c373
1bcb99b
4229600
1bcb99b
 
 
 
 
 
 
4229600
1bcb99b
4229600
1bcb99b
4229600
1bcb99b
4229600
 
1bcb99b
 
 
 
 
 
 
 
 
a72c373
 
1bcb99b
a72c373
 
1bcb99b
 
 
 
 
a72c373
 
1bcb99b
4bbd1c5
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
---
library_name: pytorch
language:
- en
tags:
- robotics
- robodojo
- vision-language-action
- meituan-robotics
---

# Meituan Robotics 0

[Technical Report](https://arxiv.org/abs/2611.08577)

A RoboDojo dual-arm vision-language-action policy with episode-first RGB and three seconds of CD-LAM visual memory. It predicts 50 actions and executes 16, using continuous gripper widths.

Inference and evaluation use the **Meituan_Robotics_0** adapter in [XPolicyLab](https://github.com/XPolicyLab/XPolicyLab/tree/main/policy/Meituan_Robotics_0). Use the matching adapter supporting the layout below.

## Model files

```text
base_vlm/                 # Qwen configuration, tokenizer and processor only
cdlam/model.pt            # Frozen visual-memory encoder
checkpoints/model.pt      # Complete VLA checkpoint, including VLM weights
config.yaml               # Inference configuration
dataset_statistics.json   # State/action normalization
```

The VLM parameters are already included in `checkpoints/model.pt`; no separate base-VLM weights or training initialization are required.

## Load

Download this repository, install the policy dependencies following the [XPolicyLab adapter README](https://github.com/XPolicyLab/XPolicyLab/blob/main/policy/Meituan_Robotics_0/README.md), and run from a RoboDojo workspace with `XPolicyLab` importable. Python 3.11 and a CUDA GPU are required.

```python
import os
from XPolicyLab.policy.Meituan_Robotics_0.model import Model

os.environ["ROBODOJO_ASSET_ROOT"] = "/path/to/downloaded/model-repository"
policy = Model({
    "bench_name": "RoboDojo", "task_name": "stack_bowls",
    "env_cfg_type": "arx_x5", "action_type": "joint",
    "ckpt_name": "mr0", "seed": 0,
})
```

Use XPolicyLab's `eval.sh` for rollouts so that observations are collected at 25 Hz for causal CD-LAM history:

```bash
# From XPolicyLab/, using your existing policy and simulator environments:
export ROBODOJO_ASSET_ROOT=/path/to/downloaded/model-repository
EVAL_ENV_TYPE=sim bash policy/Meituan_Robotics_0/eval.sh \
  RoboDojo stack_bowls mr0 arx_x5 joint 0 \
  0 1 meituan_robotics_0 robodojo
```

See [XPolicyLab](https://github.com/XPolicyLab/XPolicyLab) for installation, supported observations, deployment and evaluation instructions.

## File verification

[SHA256SUMS](SHA256SUMS) is the single SHA-256 checksum manifest for this release. It covers every repository file except the manifest itself, including weights, configuration, tokenizer/processor metadata, and documentation.

After downloading the complete repository, run from its root directory:

```bash
sha256sum -c SHA256SUMS
```