The "mind-controlling" capability has always been in mankind's fantasy. Withthe recent advancements of electroencephalograph (EEG) techniques,brain-computer interface (BCI) researchers have explored various solutions toallow individuals to perform various tasks using their minds. However, thecommercial off-the-shelf devices to run accurate EGG signal collection areusually expensive and the comparably cheaper devices can only present coarseresults, which prevents the practical application of these devices in domesticservices. To tackle this challenge, we propose and develop an end-to-endsolution that enables fine brain-robot interaction (BRI) through embeddedlearning of coarse EEG signals from the low-cost devices, namely DeepBrain, sothat people having difficulty to move, such as the elderly, can mildly commandand control a robot to perform some basic household tasks. Our contributionsare two folds: 1) We present a stacked long short term memory (Stacked LSTM)structure with specific pre-processing techniques to handle the time-dependencyof EEG signals and their classification. 2) We propose personalized design tocapture multiple features and achieve accurate recognition of individual EEGsignals by enhancing the signal interpretation of Stacked LSTM with attentionmechanism. Our real-world experiments demonstrate that the proposed end-to-endsolution with low cost can achieve satisfactory run-time speed, accuracy andenergy-efficiency.