Brain-Computer Interfaces (BCIs) decode brain activity with the aim to establish a communication channel that does not rely on muscular control. BCIs usually rely on EEG signals acquired from the subject's scalp or on electrophysiological signals from brain implants. The latter yield a superior decoding performance, but as the implant damages the cortical tissue, long-term signal stability is a concern. EEG does not require surgery but the information throughput of the BCIs is quite low and their suitability for independent daily use rather limited. Electrocorticography (ECoG) offers new perspectives for BCI. Albeit accuracy, long-term stability and independent use have already been shown, what is still lacking is the ability to control finger or hand movements from the patient's self-paced but imagined counterparts. This also summarizes the main objective of this project. But motor imagery is a skill that needs to be learnt. To meet this challenge, we will: 1) provide natural feedback to the patient by showing a hand "avatar" controlled by motor imagery; and 2) complement the decoder by what was learnt under observed movement but adjusted by the patient's ability to achieve control. This requires a new type of self-paced ECoG decoder: one that is able to assist and keep up with subject training. We will develop an iterative tensor-based decoder with automatic rank selection to maximize correctness while limiting computational and storage cost to ensure on-line performance.