Tuannm, first let me point out that your images are reversed (left on right and right on left).
I did not played seriously with the calibration part, but several problems reported on the forum were due to bad input data. In particular, you can set drawCorners to True to check if all detected chessboards are at the right position and in the same orientation in each image (seems good). If I am right, the chessboard also needs to be tilted (one side closer to the camera, with different angles) so that the perspective is detectable by the algorithm. I do not know if this is the reason, but you can maybe try to add more images with tilted chessboards and with more accentuated angles (ill-conditioned matrix means roughly that small input [corner position] errors will lead to large output [calibration data] errors, probably because the data are not distinct enough to get a significative result).
OpenCV, Python and other image processing questions