Skip to main page content
Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
, 31 (8), 3110-28

Adaptation to a Cortex-Controlled Robot Attached at the Pelvis and Engaged During Locomotion in Rats

Affiliations

Adaptation to a Cortex-Controlled Robot Attached at the Pelvis and Engaged During Locomotion in Rats

Weiguo Song et al. J Neurosci.

Abstract

Brain-machine interfaces (BMIs) should ideally show robust adaptation of the BMI across different tasks and daily activities. Most BMIs have used overpracticed tasks. Little is known about BMIs in dynamic environments. How are mechanically body-coupled BMIs integrated into ongoing rhythmic dynamics, for example, in locomotion? To examine this, we designed a novel BMI using neural discharge in the hindlimb/trunk motor cortex in rats during locomotion to control a robot attached at the pelvis. We tested neural adaptation when rats experienced (1) control locomotion, (2) "simple elastic load" (a robot load on locomotion without any BMI neural control), and (3) "BMI with elastic load" (in which the robot loaded locomotion and a BMI neural control could counter this load). Rats significantly offset applied loads with the BMI while preserving more normal pelvic height compared with load alone. Adaptation occurred over ∼100-200 step cycles in a trial. Firing rates increased in both the loaded conditions compared with baseline. Mean phases of the discharge of cells in the step cycle shifted significantly between BMI and the simple load condition. Over time, more BMI cells became positively correlated with the external force and modulated more deeply, and the network correlations of neurons on a 100 ms timescale increased. Loading alone showed none of these effects. The BMI neural changes of rate and force correlations persisted or increased over repeated trials. Our results show that rats have the capacity to use motor adaptation and motor learning to fairly rapidly engage hindlimb/trunk-coupled BMIs in their locomotion.

Figures

Figure 1.
Figure 1.
Experimental setup and training protocol. A, Rats were trained to walk on a treadmill with robot attached to the pelvis through a pelvic implanted orthosis, and neural activities were recorded from the motor cortex of the hindlimb/trunk area. The BMI could be engaged or disengaged (“switch K”) while a simple elastic load was applied to a rat. RTOS, Real-time operating system of robot. Blackrock Cerebus, Neural acquisition system. B, Training sessions were composed of three trials with 2 min baseline (BL), 2 min simple elastic load (E) trial, and 2 min BMI with elastic load trial (BMI/E) with 5–10 min resting time between trials. The elastic load on the rat and the interaction force differed in each case: in baseline there was no force, but kinematics of the pelvis were collected with the robot; in simple elastic load trials, interaction force was an elastic load (solid downward green arrow) and there was no BMI effect; in BMI with elastic load trials, interaction force was the sum of the simple elastic load (green arrow) and a neural driven BMI elastic force (solid red arrow). The neural driven force in the control was calculated from the aggregate firing rate in a 100 ms window, which modulated the stiffness of a lifting elastic field (solid red arrow) in BMI with elastic load condition. The instantaneous virtual neural driven force in baseline and simple elastic load could be calculated off-line from neural firing and kinematics (upward dotted red arrows). C, Individual neurons were detected based on the waveforms of the spikes and used in the BMI. The on-line detected spikes from five channels are shown before off-line sorting. One unit is shown before and after off-line sorted is shown below with the scatter plot of the first two tetrode principal components. Neural firing was recorded and could be correlated to kinematics or interaction forces when the rats adapted in different conditions. D, The step cycle was identified from robot tip motion py position (red), collected at 1 kHz, which corresponds to the pelvis lateral motion or step width, and this measure correlated well to video tracking as checked by toe marker from our video tracking system (blue).
Figure 2.
Figure 2.
Example of robot position patterns and forces patterns in one session in a rat (also shown in Figs. 4, 5, data). BL, Baseline; E, elastic load; BMI/E, BMI with elastic load. A–C, The average robot position (A), total interaction force (B), and neural driven force (C) in the step cycle (n = 76 steps) showed different patterns in the baseline, simple elastic load, and BMI with elastic load conditions. The equilibrium position of the load field (A, gray line) for the elastic load field in simple elastic load and BMI with elastic load conditions is shown in A. The thick dotted vertical lines correspond to the right leg swing phase in the average cycle. Colored ribbon lines shown corresponds to mean ± 2.58 SE. The rat showed more normal pelvic kinematics and less loading in BMI/E compared with E. D–G, Comparisons of mean and variance of interaction forces and neural driven forces in step cycle (n = 76 steps). D, There was a significantly decrease in the BMI with elastic load trials in mean interaction force compared with the preceding simple elastic load trials. The average total virtual interaction forces in simple elastic load (E:virtual), which were calculated by supposing the rats engaged in BMI with elastic load, were also significantly different from the real interaction forces in BMI with elastic load trial. E, The variances of the interaction force under elastic load and BMI with elastic load conditions were also significantly different and were highest in E:virtual. F, The mean virtual neural driven forces in baseline were significantly different from the mean virtual neural driven forces in simple elastic load and the mean real neural driven forces in BMI with elastic load. G, The variances of the neural driven forces were not significantly different (p > 0.05) between simple elastic load, baseline, and BMI with elastic load conditions. Statistics was applied by two-sample t test (*p < 0.01; n = 76 steps). Thus, the different offset of force and reduced variance of interaction force seen in BMI with elastic load compared with E and E:virtual was likely attributable to specific coordination patterns, as suggested in C. Error bars indicate SD.
Figure 3.
Figure 3.
Changes of force and variance of pelvic motion (Py) and mean neural firing rates across all rats and sessions. A, The rats in simple elastic load (E) showed a significantly depressed pelvic mean position, compared with baseline (BL), whereas BMI with elastic load (BMI/E) showed a height that was not significantly different from baseline but significantly different from E. The rats in BMI with elastic load normalized pelvic height. B, Mean interaction force loading in E was significantly greater than in BMI/E:real and also greater than the calculated E:virtual. C, Interaction force variance in E:virtual was greater than either BMI/E:real or E:real. D, Mean neural driven force in BMI with elastic load matched E:virtual neural driven force. Both were significantly larger than baseline (BL:virtual) neural driven force. E, E:virtual neural driven force variance was significantly greater than baseline virtual force (BL:virtual) but not BMI with elastic load (BMI:real). F, Mean neural firing rates normalized to baseline in both the first half-trial and second half-trial differed significantly in E and BMI/E compared with BL, consistent with D (neural driven forces). Both were slightly but not significantly reduced in the second half-trial compared with the first. Differences were tested by signed rank test in A–F (*p < 0.05; **p < 0.01). G, Individual cell significant rate changes obtained with paired t tests showed high percentages of cells with significant rate increases (>40%) compared with decreases (>20%) from BL to E and BL to BMI/E, but ∼25% increase and 25% decrease in comparisons from E to BMI/E. The mean and variance of the interaction forces and neural driven forces in step cycle across all sessions and rats showed similar patterns of variation and statistics as in the single trial in Figure 2 (Fig. 2D–G), except that the variance of the neural driven force also showed some differences. Neural rates showed changes consistent with these effects. Rats in BMI with elastic load normalized their pelvic height kinematics, but at the same time increased neural firing rates and reduced interaction force loads, thus differing from both simple elastic load and baseline in various ways. Error bars indicate SD.
Figure 4.
Figure 4.
The ensemble firing patterns in the step cycle in each condition in the rat and trial also displayed in Figures 2 and 5. A, The normalized firing rates of neurons shown as a thermogram and as the average of the ensemble firing rate in the step cycle, respectively. These showed different patterns under baseline (BL), simple elastic load (E), and BMI with elastic load conditions (BMI/E). The average firing rate was normalized to the firing rate peak across baseline, elastic load, and BMI with elastic load conditions. B, First half-trial in each condition shown as a thermogram. C, Second half-trial in each condition shown as a thermogram. D, E, The whole-trial average rates normalized to the whole-trial peak in each condition. Vertical dotted lines, Peak timings; dotted horizontal bars, average swing phase; solid horizontal bars, full width at 0.5. D, Whole trial. E, First half-trial. F, Second half-trial. In each half-trial, cells demonstrated different firing patterns in different adaptation fields as well as different training phases during locomotion. In the baseline condition, cells fired at specific preferred step phases in the first half and then increased phase dispersion to spread much broader in the second half-trial. In the simple elastic load condition, these cells have wider firing phase dispersion initially and then also spread in the second half-trial, although there were still clear firing phases. In the BMI with elastic load condition, there were strong firing phase preferences in both the first half-trial and the second half-trial.
Figure 5.
Figure 5.
Examples of single-cell firing pattern in the step cycle in a session (same data sets as Figs. 2, 4). A, Cells changed firing patterns in different ways (fr, firing rate in spikes/second; r, resultant vector length; θ, mean phase of the firing rate; p, p value of the nonuniformity Rayleigh test) under different conditions. The red arrow indicates the mean preferred firing phase and the resultant vector length. Cell 1 and cell 2 show common patterns of increasing and changing phase modulations. B, The firing pattern of the ensemble of all cells (n = 13) in a session. Cells showed different nonuniform discharge patterns in the different conditions (BL, 11; E, 12; BMI/E, 12). In the cell ensemble, there was a significant mean firing phase shift (nonuniform firing cells at each of the comparing conditions) between simple elastic load and either baseline or BMI with elastic load (Watson–Williams test, p < 0.05; n = 11). The red arrows indicate the mean of the ensemble preferred firing phase and the resultant vector length. C, For the same trial and population of cells (n = 13), there was significant increase in the resultant vector length during BMI with elastic load compared with either baseline or simple elastics load conditions (paired t test, p < 0.05). D, The fraction of the cycle with bursting was significantly increased during simple elastic load and BMI with elastic load condition compared with the baseline (paired t test, p < 0.05). E, The mean phase of the firing peak in the step cycle was significantly different between simple elastic load compared with either baseline or BMI with elastic load conditions (Watson–Williams t test, p < 0.05). F, Single-cell firing patterns in step cycle during the first half and the second half of trial. The blue arrows indicate the mean preferred firing phase and the resultant vector length of each cell. There was no significant difference in resultant vector length between the first half-trial and the second half-trial, but there was significant difference in the mean preferred phase of the ensemble (red arrow) between first half-trial and the second half-trial during both baseline condition (Watson Williams test, p < 0.05; n = 9) and BMI with elastic load condition (Watson Williams test, p < 0.05; n = 12). Error bars indicate SD and asterisks indicate significant difference (p < 0.05).
Figure 6.
Figure 6.
Peak firing phases in the step cycle and peak firing phase changes during adaptations in the combined data from all rats. A, Distribution of nonuniformly firing cells peak firing. Some cells preferentially fired at particular step phases under different conditions (BL, 256; E, 292; BMI/E, 293 of the total 573 cells; all with Rayleigh test, p < 0.05). Among these, 216 of 573 showed nonuniform discharging in step cycle across all conditions, whereas 96 of 573 showed uniform discharging across all conditions. We subtracted peak phases of cells which showed nonuniform discharging between conditions to examine peak phase shifts (i.e., changes in peak phase angles between conditions in B and C). Thus, 0 phase shift after subtraction indicates no change in peak phase. The numerator in the fraction number in each B and C was the number of cells shown nonuniform discharging at both conditions, and denominator was the number at either conditions. B1–B3, Differences in cell peak firing phases in first half-trial. Mean phase differences are shown by the gray arrow in each panel. In the first half-trial, there were significant changes of firing peak phase between baseline and both simple elastic load (B1; *circular t test, p < 0.05) and BMI with elastic load (B2; *p < 0.05), whereas there was no significant difference between simple elastic load and BMI with elastic load (B3; no phase shift, circular t test, p > 0.05). C1–C3, Differences in cell peak firing phases in second half-trial. Mean phase differences are shown by the gray arrow in each panel. In the second half-trial, there was no significant difference in the firing peak phases between simple elastic load and baseline (C1; no phase shift, circular t test, p > 0.05), but a significant difference persisted between BMI with elastic load and baseline (C2; *p < 0.05), and between BMI with elastic load and simple elastic load (C3; *p < 0.05). The data were consistent with a transient peak phase shift in simple elastic load condition and a longer term peak phase shift in firing in BMI with elastic load condition (*circular t test, p < 0.05), which reversed compared with the initial changes (compare B2 and C2 measured relative to baseline).
Figure 7.
Figure 7.
Correlations of individual neural firing rates with kinematic and kinetic parameters examined at the whole-population level (n = 573). A, The correlation coefficients of the firing rate of cells with both vertical force (FzBMI/E) and vertical position (PzBMI/E) were significantly increased in BMI with elastic load (red) compared with simple elastic load (FzE and PzE) (green; *kstest2, p < 0.05). Cells also tended to be more correlated with force (Fz) than with position (Pz) in both simple elastic load and BMI with elastic load conditions (*kstest2, p < 0.05). The distribution of correlation coefficients between the robot position (Pz) and firing rate was not significantly different between baseline and simple elastic load (PzBL vs PzE, kstest2, p > 0.05). B, The distribution of correlation coefficients between firing rate and measured forces or derived forces (virtual forces) under different conditions: virtual force in baseline (FBL virtual neural), virtual neural driven resulting interaction force in simple elastic load (FE virtual neural), and real total interaction force in elastic load (FE observed) and real interaction force in BMI with elastic load (FBMI/E observed). All were significantly different (*kstest2, p < 0.05). Correlations with virtual forces were much larger than actually observed. C, Changes in mean correlations with kinematic and force parameters in the BMI/E condition. In half-trial comparisons, the correlation coefficient of firing rate with force was significantly increased from the first half-trial to the second half-trial in BMI/E, whereas other kinematics parameters were either not significantly different or significantly decreased (e.g., Py) (*paired t test, p < 0.05). D, In two rats with multiple iterations of the baseline→simple elastic load→BMI with elastic load presentations in succession, the mean correlation of neurons to force was significantly altered between first and second iterations and then persisted in both rats (data for a single rat also shown in E; t test, p < 0.001). E, The correlation of neural firing to net force on the first and second iterations of presentations summarized as mean correlations in D. Note the general increase in BMI correlation over iterations. Some cells alter correlation sign between iterations, with two marked on the figure changing from positive to negative (red arrows) and four from negative to positive (blue arrows). Error bars indicate SD.
Figure 8.
Figure 8.
Force/rate amplitude modulation examined in detail. A, The firing rate of cells demonstrated linear correlation with force amplitude as seen in a typical session (same data set as in Fig. 4). The modulation strength was calculated as the linear coefficient of regressed firing rate with force amplitude (slopes of the regression, red lines). B, The complete population of neurons: The numbers of neurons with significant rate modulation with force amplitude was observed to be much larger in BMI with elastic load condition than in the simple elastic load condition. More cells were force amplitude modulated under BMI with elastic load condition (320 of 573) than under simple elastic load (216 of 573) condition. C, D, The distribution of modulation strength under simple elastic load condition (C) and under BMI with elastic load condition (D) in successive half-trials. Significantly more cells became significantly force related in the second half of the trial in BMI with elastic load condition (from 211 to 245) than were significantly force related in the first half of the trial (D) (binofit, p < 0.05). In contrast, this change was weaker under simple elastic load condition (from 126 to 142) and not significant (C) (*binomial test, p < 0.05).
Figure 9.
Figure 9.
Distribution changes of force-modulated cells between conditions and particularly in BMI. A, B, Of the 573 cells examined, ∼30% of the cells showed no modulation in either simple elastic load or BMI with elastic load conditions (gray). Thirteen percent showed force modulation only in simple elastic load (orange type I, Esig→BMI/Ensig), 30% showed force modulation only in BMI with elastic load condition (green type II, Ensig→BMI/Esig), and 30% showed modulation in both simple elastic load and BMI with elastic load condition (violet red type III, Esig→BMI/Esig). Cells could be either positively or negatively modulated. A small number of modulated cells (2%) in elastic load condition reversed their modulation strength from negative to positive, whereas most of the cells (24%) kept the modulation direction (positive to positive or negative to negative). B, Most of the positively force-modulated cells in BMI with elastic load condition came from new cells that had showed no force modulation in the simple elastic load condition. Thirteen percent of cells showed significant modulation under elastic load only (type I, Esig→BMI/Ensig); 30% of cells showed BMI with elastic load only (type II, Esig→BMI/Esig); 28% of cells showed both elastic load and BMI with elastic load (type III, Ensig→BMI/Esig). More condition specific cells were recruited during BMI with elastic load condition than during elastic load condition. C, There was significant difference in the cumulative distribution of modulation strength for force-modulated cells between simple elastic and BMI with elastic load (kstest2, p < 0.05). More cells were positively force-modulated in the BMI with elastic load condition than simple elastic load. D, There was a significant difference in the distribution of the type I and type II cells. Type I cells (drop outs of significance) were evenly distributed in modulation strength, whereas type III cells (cells newly significantly force related in BMI/E) were most likely positive-force related. E, There was no significant difference in the overall cumulative distribution of modulation strengths of type II cells (cells significantly correlated in both E in BMI/E) between the conditions of elastic field and BMI with elastic field (*significant difference with kstest2, p < 0.05). F, There was no significant modulation strength change of the individual neurons in group II between simple elastic load and BMI with elastic load.
Figure 10.
Figure 10.
Cell pair correlation coefficient and a sketch of the neural adaptation process. The cell pair correlation coefficient in half-trials across conditions (∼120 steps). A, The first half. B, The second half. The cell pair correlation matrix showed different patterns in different adaptation fields during locomotion. In the baseline condition (BL column), cell pair correlations showed a steady state across the whole trial (first half-trial vs the second half-trial). In the simple elastic load condition (E column), the cell pair correlations increased sharply in the first half-trial compared with baseline and then decreased to below the baseline condition in the second half-trial. In the first half-trial of the BMI with elastic load condition (BMI/E column), there was a moderate increase in the cell pair correlations compared with the baseline condition and the elastic load field condition, whereas in the second half-trial the cell pair correlations persisted rather than returning to the baseline condition like the simple elastic load condition. C, A simple diagram summarizes the overall neural modulation changes and differences that occurred under different field conditions and over time during the locomotion adaptation processes.

Similar articles

See all similar articles

Cited by 14 PubMed Central articles

See all "Cited by" articles

Publication types

LinkOut - more resources

Feedback