🤖 A1: A Fully Transparent Open-Source, Adaptive and Efficient Truncated Vision-Language-Action Model

---

📋 Table of Contents

• 🔧 Installation
• ⚙️ Environment Setup
• 🚀 Deploy
• 📊 Evaluation
• 🎓 Training
• 📦 Model Zoo & Datasets

---

🔧 Installation

conda create -n a1 python=3.10
conda activate a1
pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124
pip install -e .[all]
pip install --no-deps --force-reinstall git+https://github.com/moojink/dlimp_openvla
pip install -r requirements.txt 

⚙️ Environment Setup

🔐 1. Copy Environment Template

📋 cp .env.example .env.personal

✏️ 2. Configure Your Settings

Edit .env.personal with your personal settings:

# Example content
CONDA_ROOT=/path/to/conda
CONDA_ENV=a1
WANDB_ENTITY=your_entity
WANDB_PROJECT=your_project

🔒 This file is Git-ignored and won't be committed!

🔄 3. Load Environment

source .env.personal

⚠️ Security Note: .env.personal contains sensitive information (paths, API keys, etc.). Do NOT commit it to Git.

---

🚀 Deploy

Start the API server for model inference:

🖥️ bash deploy/deploy.sh --weight /put/checkpoint/here --port <port>

📋 Arguments:

✨ Example:

bash deploy/deploy.sh --weight ./model/checkpoints/pretrain --port 8000

---

📊 Evaluation

🎮 LIBERO Evaluation

📦 1. LIBERO Installation

📥 git submodule update --init robot_experiments/libero/LIBERO
📥 pip install -e robot_experiments/libero/LIBERO

🚀 2. Run Evaluation

🧪 VLABench Evaluation

VLABench evaluation requires running both a server and a client in separate terminals.

📋 Setup Steps:

🔧 Step 1: Install VLABench

pip install -r robot_experiments/vlabench/VLABench/requirements.txt
pip install -e robot_experiments/vlabench/VLABench

📥 Step 2: Download Assets (if not already downloaded)

cd robot_experiments/vlabench/VLABench
python scripts/download_assets.py --choice all

🖥️ Step 3: Start the Evaluation Server (Terminal 1)

# Load environment and start the API server
bash deploy/deploy.sh --weight <path_to_checkpoint> --port 8000

🎮 Step 4: Run the Evaluation Client (Terminal 2)

cd robot_experiments/vlabench
python eval_client.py

⚠️ Note: The server and client must run in separate terminals. The server loads the model and waits for client connections, while the client sends evaluation requests and receives results.

🏆 RoboChallenge Evaluation

RoboChallenge evaluation is executed through the run_task.py script, supporting two modes:

🦾 Supported Robot Types:

• ALOHA 🤖
• ARX5 🔧
• UR5 ⚡
• FRANKA 🦿

🧪 Mock Mode (Local Automatic Evaluation)

# 💻 Terminal 1: Deploy model
bash deploy/deploy.sh --weight <path_to_checkpoint> --port 8000

# 🎮 Terminal 2: Run mock evaluation
cd robot_experiments/RoboChallengeInference
python run_task.py \
    --task_name open_the_drawer \
    --test_type mock \
    --url http://localhost:8000

🌍 Real Mode (Official Evaluation)

# 💻 Terminal 1: Deploy model
bash deploy/deploy.sh --weight <path_to_checkpoint> --port 8000

# 🤖 Terminal 2: Run real robot evaluation
cd robot_experiments/RoboChallengeInference
python run_task.py \
    --task_name open_the_drawer \
    --test_type real \
    --url http://localhost:8000 \
    --user_token <your_token> \
    --run_id <run_id> \
    --action_nums 30

💡 Tips:

• ✅ Mock mode automatically starts the mock server, no manual startup required
• 🔑 Real mode requires valid user_token and run_id for official evaluation
• 📋 task_name must be defined in task_config.ROBO_CHALLENGE_TASKS

---

🎓 Training

🌟 Pretraining

Pretraining trains the model from scratch using large-scale VLA datasets, supporting distributed training on Slurm clusters.

📁 Configuration files:

• 📄 configs/experiments/pretrain.yaml - Pretraining experiment configuration
• 📄 configs/datasets/pretrain.yaml - Pretraining dataset configuration
• 🖥️ scripts/slurms/pretrain.sh - Pretraining script (runs on Slurm cluster)

🖥️ Slurm Cluster Training:

1️⃣ Configure Slurm submission script scripts/slurms/submit_job.sh:

• 🔧 nnodes: Number of nodes required (default 8 nodes)
• 🔧 gpus_per_node: GPUs per node (default 8)
• 🔧 partition and quotatype: Partition name and QOS type

2️⃣ Submit pretraining job:

🚀 bash scripts/slurms/submit_job.sh

💻 Single-node multi-GPU training (non-Slurm):

🚀 bash scripts/slurms/pretrain.sh

⚠️ Note: Pretraining requires significant computational resources. Distributed training on Slurm clusters is recommended. Global batch size = 128 × number of nodes.

---

📚 LIBERO Training

LIBERO training fine-tunes on simulation data, supporting single-node multi-GPU training.

📁 Configuration files:

• 📄 configs/experiments/libero_simulation.yaml - LIBERO training configuration
• 📄 configs/datasets/libero_4_tasks.yaml - LIBERO 4-task dataset configuration

🚀 Run training:

bash train_libero.sh

---

🧪 VLAbench Training

VLAbench training fine-tunes in the VLAbench simulation environment.

📁 Configuration files:

• 📄 configs/experiments/vlabench.yaml - VLAbench training configuration
• 📄 configs/datasets/vlabench.yaml - VLAbench dataset configuration

🚀 Run training:

bash train_vlabench.sh

---

🏆 RoboChallenge Training

RoboChallenge training uses the train_rc.sh script to fine-tune on specific tasks (e.g., open_the_drawer, put_cup_on_coaster).

📁 Configuration files:

🚀 Run training:

bash train_rc.sh

💡 Tip: Modify the vla_config_path variable in the script to switch between different task configurations.

---

📦 Model Zoo & Datasets

🤖 Pretrained Models

📊 Training Datasets

📁 Storage Paths:

• Model weights: Place downloaded model weights in the model/ directory
• Training data: Place downloaded datasets in the data/ directory

⚠️ Important Notes:

1. RoboMIND dataset preprocessing: Before using RoboMIND dataset, you need to run the indexing script:

   bash scripts/robomind_build_index.sh
   

2. LeRobot dataset patch for pretraining: Before pretraining, you must replace the LeRobot dataset file:

   cp a1/data/vla/lerobot_datasets_replace.py <CONDA_ENV_PATH>/lib/python3.10/site-packages/lerobot/datasets/lerobot_dataset.py
   

   Replace <CONDA_ENV_PATH> with your actual conda environment path (e.g., /path/to/conda/envs/a1)

📝 Note: Please fill in the actual download links for models and datasets in the table above.

---

🙏 Acknowledgements

This project is built upon the Molmo project. We thank the Allen Institute for AI for their excellent open-source work.

---

📚 Citation

If you find this work useful for your research, please consider citing:

@misc{zhang2026a1fullytransparentopensource,
      title={A1: A Fully Transparent Open-Source, Adaptive and Efficient Truncated Vision-Language-Action Model}, 
      author={Kaidong Zhang and Jian Zhang and Rongtao Xu and Yu Sun and Shuoshuo Xue and Youpeng Wen and Xiaoyu Guo and Minghao Guo and Weijia Liufu and Liu Zihou and Kangyi Ji and Yangsong Zhang and Jiarun Zhu and Jingzhi Liu and Zihang Li and Ruiyi Chen and Meng Cao and Jingming Zhang and Shen Zhao and Xiaojun Chang and Feng Zheng and Ivan Laptev and Xiaodan Liang},
      year={2026},
      eprint={2604.05672},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2604.05672}, 
}