cronusvla_7B_libero_spatial

Model Card for CronusVLA-Libero

• checkpoints/*.pt: Complete model checkpoint of CronusVLA-7B for direct evaluation on Libero-spatial. It only conditions on the textual instruction and a third-person image (no wrist-view image and no state).

• *.txt: the evaluation results with 3 random seeds

• dataset_statistics.json: the normalization statistics for the dataset

If you want to evaluate or further finetune with this checkpoint, please refer to CronusVLA for more details.

Evaluation:

#!/bin/bash

CUDA_DEVICES=(0 1 2 3 4 5 6 7)  # can be modified according to demand

# pre-trained model storage directory
CHECKPOINT_DIR=./checkpoints

CHECKPOINTS=($(ls "$CHECKPOINT_DIR"/*.pt | sort))

# CUDA devices number
NUM_CUDA_DEVICES=${#CUDA_DEVICES[@]}
INDEX=0

for CHECKPOINT in "${CHECKPOINTS[@]}"; do
    CUDA_DEVICE=${CUDA_DEVICES[$((INDEX % NUM_CUDA_DEVICES))]}  # polling to allocate GPU
    echo "Running on GPU $CUDA_DEVICE with checkpoint $CHECKPOINT"
    
    CUDA_VISIBLE_DEVICES=$CUDA_DEVICE python experiments/Libero/robot/libero/run_libero_eval.py \
        --model_family cronus \
        --pretrained_checkpoint "$CHECKPOINT" \
        --task_suite_name libero_spatial \
        --center_crop True \
        --use_wrist_image False &
    
    sleep 2
    ((INDEX++))
done

wait

📚 Citation

If you find this model useful, please cite our work:

@article{li2025cronusvla,
  title={CronusVLA: Transferring Latent Motion Across Time for Multi-Frame Prediction in Manipulation},
  author={Li, Hao and Yang, Shuai and Chen, Yilun and Tian, Yang and Yang, Xiaoda and Chen, Xinyi and Wang, Hanqing and Wang, Tai and Zhao, Feng and Lin, Dahua and others},
  journal={arXiv preprint arXiv:2506.19816},
  year={2025}
}