2。3。3Design of the ErRP detector

The visual stimulus, shown in Fig。 3e, trains the users to timely localize the occurrence of error through visual feed- back。 Here, the stimuli begins with a 30-s recording of baseline EEG followed by a repetition of trials, where the instructions to the user are given。 The timing scheme of an inpidual trial is as follows: a fixation cross for 1 s, fol- lowed by an error detection task for 2 s and a blank screen for 2 s。 As shown in Fig。 3e, during the error detection task, a simulated robot link moves toward the target。 If the end point of the moving link crosses the target, an error is iden- tified by the user。 During training over three sessions, 50 trials have errors incorporated in them and 200 trials have correct responses。 The task of the user is to detect the error trials。

Similar to P300 components, an ErRP waveform is a time-locked event occurring when the subject perceives an error during a task。 The ErRP is characterized by a nega- tive peak at around 150 ms followed by a positive peak at around 200–500 ms, after an error response is detected   at a frequency range of 1–10 Hz [7, 22]。 This feature of the ErRP waveforms is used for error detection in our study。

Here too, a filter of the same specification as the one used for P300 signal detection is used to filter the incoming signal from Fz electrode for a frequency range of 1–10 Hz。 Similar to P300 component detection, the incoming EEG signals are averaged and used for training a linear kernel- SVM classifier。

The testing session is similar to training session with an additional feedback period of 2 s after every cue。 Here too, an additional feedback period of 2 s is incorporated after the instruction cue in the online testing stimuli, shown in Fig。 2f。 The test stimuli comprises 20 trials have error responses and 80 trials have correct responses。

The construction of the feature vectors and testing of the ErRP classifier are similar to the one used for P300 detec- tion。 When the ErRP detector yields an output y = 1, it

means an error has occurred。 Next, the detector notes the

previous state of the system。 If the MI detector was acti- vated prior to the activation of the ErRP detector (occur- rence of directional error), the system stops the movement of the robot arm and sends it back to its previous position。 If the P300 detector was activated prior to the activation of the ErRP detector (occurrence of positional error), the system realigns the robot arm to the target position by an

Table 2  Classification accuracy

(C。A。 in %) and information transfer rate (ITR in bits/min) of five subjects for training and online testing of the motor imagery (MI), P300 and ErRP detectors

Best results marked in bold

Detectors Subject ID Training Online testing

C。A。 TPR FPR C。A。 ITR

offset。 This step is the main difference between the ErRP detector and the P300 detector。 Now if y = 0, then the sys- tem continues moving the robot arm。

2。4Real-time robot arm controller design

In the present application, the control signal is required for a very small duration in the order of 300 ms for the correction of the positional error。 Here, the controller logic combines the outputs of all three decoders to perform a movement task。 Two different controllers have been designed for the proposed scheme。 The first controller acts like a typical on– off controller, and it is deactivated with the recognition of P300 signal。 The second controller is activated if an ErRP signal is detected。 Here, we  generate a control  command to turn the robot arm in the reverse direction of its previous movement by a small offset。 The controller functions, devel- oped using the above philosophy, are given below。