CEED-VLA: Consistency Vision-Language-Action Model with Early-Exit Decoding

Wenxuan Song¹*, Jiayi Chen¹*, Pengxiang Ding²˒³†, Yuxin Huang¹, Han Zhao²˒³, Donglin Wang², Haoang Li¹‡

¹ IRPN Lab, HKUST(GZ)
² MiLab, Westlake University
³ Zhejiang University

* Equal Contribution  † Project Leader  ‡ Corresponding Author

Overview | Installation | Get Checkpoint | Evaluation | Results

---

Overview

We now only open-source the consistent-training-free version of CEED-VLA base on OpenVLA.
The consistent training code , jacobi trajectory generating code and LLaVA-VLA version will be release later.

Installation

Note: If you already have the OpenVLA environment installed, you can skip this section.

1. Install OpenVLA.

# Create and activate conda environment
conda create -n openvla python=3.10 -y
conda activate openvla

# Install PyTorch. Below is a sample command to do this, but you should check the following link
# to find installation instructions that are specific to your compute platform:
# https://pytorch.org/get-started/locally/
conda install pytorch torchvision torchaudio pytorch-cuda=12.4 -c pytorch -c nvidia -y  # UPDATE ME!

# Clone and install the openvla repo
git clone https://github.com/IRPN-EAI/CEED-VLA.git
cd openvla
pip install -e .

# Install Flash Attention 2 for training (https://github.com/Dao-AILab/flash-attention)
#   =>> If you run into difficulty, try `pip cache remove flash_attn` first
pip install packaging ninja
ninja --version; echo $?  # Verify Ninja --> should return exit code "0"
pip install "flash-attn==2.5.5" --no-build-isolation

2. Install Libero

Clone and install the LIBERO repo:

git clone https://github.com/Lifelong-Robot-Learning/LIBERO.git
cd LIBERO
pip install -e .

Additionally, install other required packages:

cd openvla
pip install -r experiments/robot/libero/libero_requirements.txt

Download dataset

git clone [email protected]:datasets/openvla/modified_libero_rlds

Get the checkpoint

Download checkpoint

We provide a pretrained checkpoint of OpenVLA with action chunk = 3.

git lfs install

git clone https://huggingface.co/chenpyyy/openvla-ac

(option) Train a checkpoint with action chunk=3 by youself

Note: Before running the finetuning script, please download OpenVLA from Hugging Face and place it in ./model_ckpt/openvla-7b. Then, replace the following files in the downloaded model with those in our repo:

Replace:
  ./model_ckpt/openvla-7b/configuration_prismatic.py
  ./model_ckpt/openvla-7b/modeling_prismatic.py
  ./model_ckpt/openvla-7b/processing_prismatic.py

With:
  ./prismatic/extern/hf/configuration_prismatic.py
  ./prismatic/extern/hf/modeling_prismatic.py
  ./prismatic/extern/hf/processing_prismatic.py

train!

#!/bin/bash

export DATA_ROOT_DIR="/hpc2hdd/home/haoangli/wenxuan/project/openvla-consitent/modified_libero_rlds/"  #libero_10_no_noops path
export RUN_ROOT_DIR="./run/libero_10_ac3" # save train result path                 
export ADAPTER_TMP_DIR="./run/libero_10_ac3/adapter_weights"       

# 创建必要的目录
mkdir -p $RUN_ROOT_DIR
mkdir -p $ADAPTER_TMP_DIR
# export WANDB_MODE=offline
# export CUDA_VISIBLE_DEVICES=0,1

# lora fitune
torchrun --standalone --nnodes 1 --nproc-per-node 8 vla-scripts/finetune.py \
  --vla_path "./model_ckpt/openvla-7b" \
  --data_root_dir $DATA_ROOT_DIR \
  --dataset_name libero_10_no_noops \
  --run_root_dir $RUN_ROOT_DIR \
  --adapter_tmp_dir $ADAPTER_TMP_DIR \
  --lora_rank 32 \
  --batch_size 16 \
  --grad_accumulation_steps 1 \
  --learning_rate 5e-4 \
  --image_aug True \
  --save_steps 10000 \
  --wandb_project openvla-finetune-ac3 \
  --window_size 1 \
  --future_action_window_size 2 \
  --max_steps 100000 \
  --save_latest_checkpoint_only False \
  --num_images_in_input 1

LIBERO Evaluations

Results of OpenVLA in LIBERO-LONG Environment on H100(Speed in tokens/s)

Decoding Method	Action Chunk	Success Rate	Min Speed	Max Speed	Avg. Speed
AR	1	53.0	48.1	55.3	51.0
AR	3	60.4	49.3	60.4	54.3
Jacobi (ours)	3	62.4	62.9	163.1	85.0

Each success rate is the 500 rollouts each (10 tasks x 50 rollouts per task).

Run Evaluation Script

bash ./vla-scripts/eval/run_eval_libero_10_jacobi_kv.sh

run_eval_libero_10_jacobi_kv.sh

#!/bin/bash

CUDA_VISIBLE_DEVICES=0 python ./experiments/robot/libero/run_libero_eval_jacobi_kv.py\
  --model_family openvla \
  --pretrained_checkpoint ./ckpt/openvla-7b+libero_10_no_noops+b16+lr-0.0005+lora-r32+dropout-0.0--image_aug-ac3-80000-ckpt \
  --task_suite_name libero_10 \
  --center_crop True \
  --action_chunk 3 \

Cite

@article{song2025ceed,
  title={CEED-VLA: Consistency Vision-Language-Action Model with Early-Exit Decoding},
  author={Song, Wenxuan and Chen, Jiayi and Ding, Pengxiang and Huang, Yuxin and Zhao, Han and Wang, Donglin and Li, Haoang},
  journal={arXiv preprint arXiv:2506.13725},
  year={2025}
}