Efficient Real Time control of a robot simulator with brain waves

[Rolando]

Authors : R. Grave de Peralta, Q. Noirhomme, G. Vanacker* , M. Nuttin*and S. Gonzalez Andino.

Electrical Neuroimaging Group, Neurology Dept., (Electrical NeuroImaging Group - Geneva)
Geneva University Hospital,Switzerland.
*Department of Mechanical Engineering, KUL Leuven, Belgium

Abstract

The main goal of this work is the development of a Brain Computer Interface able to satisfy the following constraints:

1) Non-invasiveness, i.e., based on the use of the EEG or measures derived from it.
2) High transfer rate: Able to interact with the robot at least every ½ second.
3) Minimal training requirements: Avoid the long training periods, reusing the classifiers obtained in previous sessions.
4) Able to function in real life conditions (environmental noise), i.e., allowing the subject to control de robot while other people talk with him as in a diary life.
5) Requirement of a low level of artificial intelligence to reduce the price of the system.
6) Able to deal with several classes (≈6).

On this poster we present a solution which respond satisfactory to the first 5 points while solving only partially constraint 5 since only three classes have being properly identified.

In particular the Geneva Brain Computer Interface (G-BCI) described here is based on spatial visual attention and the so called steady state visual evoked potentials. While the basic experimental design resembles those previously described elsewhere [1] and [2], the main difference remains the feature selection algorithm described in [3] called the discriminative power (DP). The DP estimates, for each feature, the number of true positive (TP) given that the number of false positive is zero. Using this measure we build a linear classifier based on the Proximal SVM approach (PSVM). Finally an heuristic filtering strategy aiming at suppressing false positive is added to the output of the classifier. For now on the term classifier will refre to both, the PSVM together with the filtering strategy of the output scores. As for the features we use the oscillatory activity in the EEG extracted with a simple FFT algorithm. In summary the brain computer interface presented here is based on very fast algorithms that allow an efficient online implementation. To illustrate it we control in real time a robot simulator in different scenarios. In addition the robot simulator can be adjusted to use different artificial intelligence levels allowing also to asses the need for complex obstacle avoidance algorithms.

The data from two healthy subjects participating in 4 sessions each one, were used to evaluate theoretically the methodology proposed here. The results for both subjects were very similar. Using two visual stimulus (corresponding to Right or Left movements of the robot), we have been able to correctly identify 100% of the stimulus in ½ second and no less than 95% in ¼ second using a 10 fold cross validation procedure. Since the quarter of second is probably not useful for the task presented here, we will consider in the following classifications based only on half a second. This corresponds to 120 decisions per minute, that using the method proposed in [4] yields a maximum transfer rate of 120 bits/minute. In addition similar (about 99%) classification rates were obtained using classifiers build on different sessions, it means that at, least theoretically, we need to build the classifier only once.

In practice we have three classes corresponding to Left, Right or Continue to move, to transmit to the robot. Using a sliding window on the training data with 3 classes, we estimated a (very pessimistic) lower bound to the correct classification rate of 87%. This value corresponds to a bit rate of 107.7 bits/min that surpass most of the methods presented so far based on visual attention or other strategies.

These values are compatible with experimental results during the real time control of a robot simulator by one of the subjects, using a classifier stored from a training session of a previous day. This confirms that for practical purposes, the classifier can be also computed in advance. In addition the trajectory described by the robot denotes the (almost) inexistence of false positives obtained, unfortunately, at the price of certain rigidity of the “steering wheel”.

Regarding the environmental noise, the subjects refer very low disturbance even if people are talking to them, that means, the level of attention allow the subjects to change their attention from the robot to the speakers without affecting the robot control. However, if the subjects try to talk or move, false positives might appear, then, at current stage of development, we recommend the subject to refrain from talking while transiting by a narrow lane.

As for the level of artificial intelligence required, the result with two scenarios of variable complexity, show that it could reduced to an anti-collision systems which might be less expensive than a full obstacle avoidance system. All these practical results can be appreciated in the films made during the experiments and will be available on demand on our computers during the meeting.

The results of this work and the online demonstration (available soon at Electrical NeuroImaging Group - Geneva) show that visual spatial attention is an interesting alternative to build robust Human Machine Interfaces with high transfer rates and also that additional work is needed to increase the number of classes to the minimum (6) required by our constraints.

Acknowledgments: Research funded by European projects MAIA BACS and Swiss NCCR- IM2.

References:

[1] M. Cheng, D. Xu, X. Gao, S. Gao, (2002) IEEE Trans. Biom. Eng.
[2] S.P. Kelly, E.C. Lalor, R. B. Reilly & J.J. Foxe, (2005). IEEE Trans. Neural System & Rehab. Eng. 13,2, 172-178.
[3] S. Gonzalez Andino, R. Grave de Peralta, G, Thut. J. Millan, P. Morier, T. Landis (2006). Neuroimage. 32,1,170-179.
[4] , J. R. Wolpaw, H. Ramoser, D. J. McFarland, & g. Pfurtscheller, (1998). IEEE Trans. Rehab. Eng., 6, 3,326-333.