1.赛事相关信息

• 点击查看

2.赛题分析

• 软体对GPU要求较高，环境配置复杂，选择刚体环境
• 先以模仿学习/强化学习的刚体环境为基础，后期再考虑无限制刚体环境
• 部分任务（如将物块移动到指定位置），存在相机之外的额外输入：目标位置，rgbd图像难以将其作为可视化输入（超出了相机范围），所以选择点云
• 由于机械臂操作策略只关心末端执行器的运动，所以控制执行器做出动作，控制模式pd_ee_delta_pose，动作控制的是末端执行器的每一步运动

3.演示转换

3.1.演示重现

• 使用maniskill2提供的脚本：可以交互查看不同相机拍摄的内容，赛题里面只能使用两个相机（手上相机hand_camera、机械臂斜上方的相机base_camera），视野范围都很小，并且存在遮挡：

python tools/replay_trajectory.py --traj-path demos/rigid_body_envs/PickCube-v0/PickCube-v0.h5 --vis

3.2.控制模式转换

• 提供的演示示例需要进行控制模式的转换：使用maniskill2提供的脚本，10线程转换， i7-7700HQ CPU，大概五分钟，五分钟之内电脑卡的别的什么都干不了：

python tools/replay_trajectory.py --traj-path demos/rigid_body_envs/PickCube-v0/PickCube-v0.h5 \
  --save-traj --target-control-mode pd_ee_delta_pose  --num-procs 10

3.3.点云可视化

• 使用matplotlib散点图可视化

import gym
import mani_skill2.envs
import numpy as np
import open3d as o3d
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D





#obs_mode = pointcloud, rgbd, state_dict, state
env = gym.make("PickCube-v0", obs_mode="pointcloud", control_mode="pd_ee_delta_pose")


env.seed(0)  
obs = env.reset()
done = False
fig = plt.figure()
i = 0
while not done:
    action = np.array([0.1,0.1,0.1,0,0,0,1])
    obs, reward, done, info = env.step(action)
    point = np.array(obs['pointcloud']['xyzw'])
    rgb = np.array(obs['pointcloud']['rgb'])
    print(point.shape)#(32768,4)
    print(rgb.shape)#(32768,3)

    
    i += 1
    if i == 10:
        ax = fig.add_subplot(projection='3d')

        #point_xyz = point[:,[0,1,2]]
        
        ax.scatter(point[:,[0]], point[:,[1]], point[:,[2]], zdir="z",facecolors=rgb/255, marker="o", s=40)
        #ax.scatter(xs, ys, zs=zs2, zdir="z", facecolors=rgb, marker="^", s=40)

        ax.set(xlabel="X", ylabel="Y", zlabel="Z")
        plt.show()


    
    env.render()  # a display is required to render
env.close()

• 效果：
  
  

3.4.示例点云转换

• 使用maniskill2_learn脚本：

python tools/convert_state.py --env-name PickCube-v0 --num-procs 1 \
--traj-name /home/jiangyvhang/ManiSkill2/demos/rigid_body_envs/PickCube-v0/PickCube-v0.none.pd_ee_delta_pose.h5 \
--json-name /home/jiangyvhang/ManiSkill2/demos/rigid_body_envs/PickCube-v0/PickCube-v0.none.pd_ee_delta_pose.json \
--output-name /home/jiangyvhang/ManiSkill2/demos/rigid_body_envs/PickCube-v0/PickCube-v0.trajectory.none.pd_ee_delta_pose_pcd.h5 \
--control-mode pd_ee_delta_pose \
--max-num-traj  6 \
--obs-mode pointcloud \
--n-points 1200 \
--obs-frame base \
--reward-mode dense \
--n-goal-points 50 \
--obs-frame base \
--render \

• 参数解读：

--env-name：环境名字 \
--num-procs ：线程数，越多转换越快，但启动多线程会比较慢，1 \
--traj-name .h5示例文件路径 \
--json-name .json示例文件路径 \
--output-name 输出的转换后的.h5文件路径 \
--control-mode ：控制模式，pd_ee_delta_pose \
--max-num-traj  ：转换的路径个数，小一点，更快完成转换，快速验证，6 \
--obs-mode ：观测方式，pointcloud \
--n-points ：点云点数，1200 \
--obs-frame ：观测点云基准，base \
--reward-mode：选择稠密或稀疏奖赏 dense：稠密，sparse：稀疏 \
--n-goal-points ：目标位置处点云数量，50 \
--render ：演示，会降低转换速度，并且只能开启一个线程\

• 点云可视化：

import h5py
import numpy as np

file =  '/home/jiangyvhang/ManiSkill2/demos/rigid_body_envs/PickCube-v0/PickCube-v0.trajectory.none.pd_ee_delta_pose_pcd.h5'
h5file = h5py.File(file, "r")
print(np.array(h5file['traj_0']))
for key in h5file['traj_0'].keys():
    #print(h5file['traj_0'][key])
    '''
    <HDF5 dataset "dict_str_actions": shape (88, 7), type "<f4">
<HDF5 dataset "dict_str_dones": shape (88, 1), type "|b1">
<HDF5 dataset "dict_str_episode_dones": shape (88, 1), type "|b1">
<HDF5 dataset "dict_str_is_truncated": shape (88, 1), type "|b1">
<HDF5 group "/traj_0/dict_str_obs" (5 members)>
<HDF5 dataset "dict_str_rewards": shape (88, 1), type "<f4">
<HDF5 dataset "dict_str_worker_indices": shape (88, 1), type "<i4">
    '''
    #print( np.array( h5file['traj_0'][key] ) )
    pass

'''
for key in h5file['traj_0']['dict_str_obs']:
    print(h5file['traj_0']['dict_str_obs'][key])
    print( np.array( h5file['traj_0']['dict_str_obs'][key] ) )
'''
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
fig = plt.figure()
ax = fig.add_subplot(projection='3d')

point = np.array( h5file['traj_0']['dict_str_obs']['dict_str_xyz'] )
rgb = np.array( h5file['traj_0']['dict_str_obs']['dict_str_rgb'] )

i = 80#查看第i帧
ax.scatter(point[[i],:,[0]], point[[i],:,[1]], point[[i],:,[2]], zdir="z",facecolors=np.squeeze(rgb[[i],:],0), marker=".", s=40)


ax.set(xlabel="X", ylabel="Y", zlabel="Z")
plt.show()

效果：

i = 0

i = 10

i = 80

• 调整参数，重新转换（需要删除上一步输出的文件，或更改输出的文件名）：

python tools/convert_state.py --env-name PickCube-v0 --num-procs 1 \
--traj-name /home/jiangyvhang/ManiSkill2/demos/rigid_body_envs/PickCube-v0/PickCube-v0.none.pd_ee_delta_pose.h5 \
--json-name /home/jiangyvhang/ManiSkill2/demos/rigid_body_envs/PickCube-v0/PickCube-v0.none.pd_ee_delta_pose.json \
--output-name /home/jiangyvhang/ManiSkill2/demos/rigid_body_envs/PickCube-v0/PickCube-v0.trajectory.none.pd_ee_delta_pose_pcd.h5 \
--control-mode pd_ee_delta_pose \
--max-num-traj  2 \
--obs-mode pointcloud \
--n-points 10000 \
--obs-frame world \
--reward-mode dense \
--n-goal-points 50 \
--render \

• 输出：

Obs mode: pointcloud; Control mode: pd_ee_delta_pose
Obs frame: world; n_points: 10000; n_goal_points: 50
Reset kwargs for the current trajectory: {'seed': 0}
Convert Trajectory: completed 1 / 2; this trajectory has length 88
Reset kwargs for the current trajectory: {'seed': 1}
Convert Trajectory: completed 2 / 2; this trajectory has length 101
Finish using /tmp/0.h5
maniskill2_learn - (record_utils.py:280) - INFO - 2022-12-02,16:31:19 - Total number of trajectories 2
Finish merging files to /home/jiangyvhang/ManiSkill2/demos/rigid_body_envs/PickCube-v0/PickCube-v0.trajectory.none.pd_ee_delta_pose_pcd.h5

• 效果：

i = 0

i = 10

i = 80