At hybrid analog-digital (HAD) transceiver, an improved HAD rotational invariance techniques (ESPRIT), called I-HAD-ESPRIT, is proposed to measure the direction of arrival (DOA) of desired user, where the phase ambiguity due to HAD structure is addressed successfully. Subsequently, a machine-learning (ML) framework is proposed to improve the precision of measuring DOA. In the training stage, the HAD transceiver works as a receiver and repeatedly measures the values of DOA via I-HAD-ESPRIT to form a slightly large training data set (TDS) of DOA. From TDS, we find that the probability density function (PDF) of DOA measurement error (DOAME) is approximated as a Gaussian distribution by the histogram method in ML. This TDS is used to learn the mean of DOA and the variance of DOAME, which are utilized to infer their values and improve their precisions in the real-time stage. The HAD transceiver rapidly measures the real-time value of DOA some times to form a relatively small real-time set (RTS), which is used to learn the real-time mean and variance of DOA/ DOAME. Then, three weight combiners are proposed to combine the-maximum-likelihood-learning outputs of TDS and RTS. Their weight factors depend intimately on the numbers of elements in TDS and RTS, and signal-to-noise ratios during the two stages. Using the mean and variance of DOA/DOAME, their PDFs can be given directly, and we propose a robust beamformer for directional modulation (DM) transmitter with HAD by fully exploiting the PDF of DOA/DOAME, especially a robust analog beamformer on RF chain. Simulation results show that: 1) The proposed I-HAD-ESPRIT can achieve the HAD CRLB; 2) The proposed ML framework performs much better than the corresponding real-time one without training stage, 3) The proposed robust DM transmitter can perform better than the corresponding non-robust ones in terms of bit error rate and secrecy rate.