README.md

metadata

library_name: pytorch
language:
  - en
tags:
  - robotics
  - robodojo
  - vision-language-action
  - meituan-robotics

Meituan Robotics 0

Technical Report

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. Use the matching adapter supporting the layout below.

Model files

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, and run from a RoboDojo workspace with XPolicyLab importable. Python 3.11 and a CUDA GPU are required.

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:

# 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 for installation, supported observations, deployment and evaluation instructions.

File verification

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:

sha256sum -c SHA256SUMS