Views
No views yet
If you want to evaluate or further finetune with this checkpoint, please refer to CronusVLA for more details.
1#!/bin/bash
2
3CUDA_DEVICES=(0 1 2 3 4 5 6 7) # can be modified according to demand
4
5# pre-trained model storage directory
6CHECKPOINT_DIR=./checkpoints
7
8CHECKPOINTS=($(ls "$CHECKPOINT_DIR"/*.pt | sort))
9
10# CUDA devices number
11NUM_CUDA_DEVICES=${#CUDA_DEVICES[@]}
12INDEX=0
13
14for CHECKPOINT in "${CHECKPOINTS[@]}"; do
15 CUDA_DEVICE=${CUDA_DEVICES[$((INDEX % NUM_CUDA_DEVICES))]} # polling to allocate GPU
16 echo "Running on GPU $CUDA_DEVICE with checkpoint $CHECKPOINT"
17
18 CUDA_VISIBLE_DEVICES=$CUDA_DEVICE python experiments/Libero/robot/libero/run_libero_eval.py \
19 --model_family cronus \
20 --pretrained_checkpoint "$CHECKPOINT" \
21 --task_suite_name libero_object \
22 --center_crop True \
23 --use_wrist_image False &
24
25 sleep 2
26 ((INDEX++))
27done
28
29wait
301@article{li2025cronusvla,
2 title={CronusVLA: Transferring Latent Motion Across Time for Multi-Frame Prediction in Manipulation},
3 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},
4 journal={arXiv preprint arXiv:2506.19816},
5 year={2025}
6}