You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

A1 — RoboDojo joint-delta VLA (step88000)

A1 是基于 AllenAI Molmo 的视觉-语言-动作(VLA)模型。本仓库提供在 RoboTwin 3.0 / RoboDojo(ARX5 双臂)上训练的 joint-delta checkpoint,以及一套开箱即用的部署代码(XPolicyLab socket server),下载后即可起服务、被仿真器调用。

A1 is a Molmo-based Vision-Language-Action model. This repo ships a RoboDojo (ARX5 dual-arm) joint-delta checkpoint plus a ready-to-run serving stack (XPolicyLab socket server).

仓库内容 / Contents

路径	说明
`checkpoint/step88000-unsharded/model.pt`	模型权重(~32GB,unsharded)
`checkpoint/step88000-unsharded/config.yaml`	该 checkpoint 的训练配置
`assets/robodojo_joint_delta.json`	反归一化统计(`bounds_q99`),与该 checkpoint 配套
`code/XPolicyLab_a1_robodojo_full.tar.gz`	完整 XPolicyLab 部署代码(含 vendor 的 delta-capable A1、含 `.git`)

模型规格 / Model spec

机器人 / robot: ARX5 双臂 (dual-arm)
动作 / action: 14 维关节 [L_arm(6), L_grip(1), R_arm(6), R_grip(1)]
joint-delta: 双臂关节预测的是 delta,夹爪是绝对值 → delta_mask = "6,-1,6,-1"
归一化 / normalization: bounds_q99
chunk 长度 / action chunk: 50,序列长度 / sequence length: 600

快速使用 / Quick start

# 1) 下载本仓库(用 hf-mirror 更快;国内环境关掉代理)
pip install -U "huggingface_hub[cli]"
HF_ENDPOINT=https://hf-mirror.com \
  hf download D-eon/a1-robodojo-joint-delta --local-dir ./a1_robodojo

# 2) 解压部署代码
cd a1_robodojo
tar -xzf code/XPolicyLab_a1_robodojo_full.tar.gz      # 得到 XPolicyLab/

# 3) 准备 conda 环境(参考 XPolicyLab/policy/A1/README.md 与 A1 安装说明)
#    需要 torch==2.6.0 / transformers<5 等;环境名假设为 a1

# 4) 起 A1 策略服务(指向下载的 checkpoint 目录)
MODEL_PATH="$PWD/checkpoint/step88000-unsharded" \
  bash XPolicyLab/policy/A1/serve_a1.sh
#   可选: PORT=32180 GPU=0 CONDA_ENV=a1 ...
#   norm stats 默认用 XPolicyLab/policy/A1/assets/robodojo_joint_delta.json(已随代码包提供)

服务以端口进入 LISTEN 为就绪标志(checkpoint ~32GB,加载需几分钟)。详见解压后的 XPolicyLab/policy/A1/README.md(「独立部署 / 开环验证」一节)。

开环验证 / Open-loop check

python XPolicyLab/policy/A1/test_openloop_client.py \
  --episode /abs/.../RoboDojo/<task>/arx_x5/data/episode_0000007.hdf5 \
  --host 127.0.0.1 --port 32180 --chunk 50 --max_frames 400

在 arrange_largest_number 上整体 L1 ≈ 0.02–0.03(关节空间绝对动作)属正常范围。

关键对齐点 / Notes

delta 还原在 server 端完成(policy/A1/model.py):按 delta_mask 把原始 state 加回臂关节。
输入图像保留原分辨率(交给模型自带预处理),相机顺序 cam_head / cam_left_wrist / cam_right_wrist,需 RGB。
obs 必须带真实 instruction,否则精度明显下降。

Downloads last month: -; Downloads are not tracked for this model. How to track

Video Preview

Robotics