机器人持续学习基准LIBERO系列9——数据集轨迹查看

最新推荐文章于 2025-06-10 11:41:16 发布

阿航626

最新推荐文章于 2025-06-10 11:41:16 发布

阅读量1.6k

点赞数 12

分类专栏：机器人操作持续学习论文 LIBERO 文章标签：机器人 LIBERO libero robosuite 机器人学习持续学习多任务学习

本文链接：https://blog.youkuaiyun.com/weixin_53610475/article/details/136107854

版权

机器人操作持续学习论文同时被 2 个专栏收录

18 篇文章

订阅专栏

LIBERO

9 篇文章

订阅专栏

0.前置

1.获取轨迹文件路径有关包

from libero.libero import benchmark, get_libero_path
import os

2.初始化类和任务

用来获取轨迹文件文件名

benchmark_instance = benchmark.get_benchmark_dict()["libero_10"]()
num_tasks = benchmark_instance.get_num_tasks()

3.生成轨迹文件路径

datasets_default_path = get_libero_path("datasets")
print(datasets_default_path)
demo_files = [os.path.join(datasets_default_path, benchmark_instance.get_task_demonstration(i)) \
              for i in range(num_tasks)]
print(demo_files)

demo_files是由hdf5文件路径组成的列表

4.重放轨迹有关包

import h5py
from libero.libero.utils.dataset_utils import get_dataset_info
from IPython.display import HTML
import imageio

5.轨迹hdf5文件结构

demo_0:n步
- actions:(n,7)
- dones:(n,)
- obs
  - agentview_rgb:(n,128,128,3)
  - ee_ori:(n,3)
  - ee_pos:(n,3)
  - ee_states:(n,6)
  - eye_in_hand_rgb:(n,128,128,3)
  - gripper_states:(n,2)
  - joint_states:(n,7)
- rewards:(n,)
- robot_states:(329,47)
demo_1
…

6.使用手外相机生成重放视频

example_demo_file = demo_files[9]
# Print the dataset info. We have a standalone script for doing the same thing available at `scripts/get_dataset_info.py`
get_dataset_info(example_demo_file)

with h5py.File(example_demo_file, "r") as f:
    images = f["data/demo_0/obs/agentview_rgb"][()]

video_writer = imageio.get_writer("output.mp4", fps=60)
for image in images:
    video_writer.append_data(image[::-1])
video_writer.close()

HTML("""
    <video width="640" height="480" controls>
        <source src="output.mp4" type="video/mp4">
    </video>
    <script>
        var video = document.getElementsByTagName('video')[0];
        video.playbackRate = 2.0; // Increase the playback speed to 2x
        </script>    
""")