我假设你对李群和李代数有很好的理解。如果使用se（3）作为变量，通过计算损失将其转化为SE（3），则问题是无约束的，易于通过Scipy优化。那么，你的主要困难是什么？
要将se（3）（李代数）转换为SE（3）（李群），您也可以使用以下代码：

from scipy.spatial.transform import Rotation
import numpy as np

def skew(x:np.ndarray):
    return np.array([[0,-x[2],x[1]],
                     [x[2],0,-x[0]],
                     [-x[1],x[0],0]])
def computeV(rvec:np.ndarray):
    theta = np.linalg.norm(rvec)
    skew_rvec = skew(rvec)
    skew_rvec2 = skew_rvec * skew_rvec
    V = np.eye(3) + (1 - np.cos(theta))/theta**2 * skew(rvec) + (theta - np.sin(theta))/theta**3 * skew_rvec2
    return V
    
def toVec(SE3:np.ndarray):
    """SE3 Matrix to rvec and tvec

    Args:
        SE3 (np.ndarray): 4x4 `np.ndarray`

    Returns:
        rvec, tvec: `np.ndarray`
    """
    R = Rotation.from_matrix(SE3[:3,:3])
    rvec = R.as_rotvec()
    V = computeV(rvec)
    tvec = np.linalg.inv(V) @ SE3[:3,3]
    return rvec, tvec

def toMat(rvec:np.ndarray, tvec:np.ndarray):
    """rvec and tvec to SE3 Matrix

    Args:
        rvec (`np.ndarray`): 1x3 rotation vector\n
        tvec (`np.ndarray`): 1x3 translation vector
        
    Returns:
        SE3: 4x4 `np.ndarray`
    """
    R = Rotation.from_rotvec(rvec)
    V = computeV(rvec)
    mat = np.eye(4)
    mat[:3,:3] = R.as_matrix()
    mat[:3,3] = V @ tvec
    return mat

验证：

a = np.random.rand(6)
    print("se3: {}".format(a))
    SE3 = toMat(a[:3],a[3:])
    print("SE3:{}".format(SE3))
    print("rvec:{}, tvec:{}".format(*toVec(SE3)))