Figure 5.

RNN decoders enable two-cursor control and outperform simple linear decoders. (a) Median target acquisition time and angular errors are shown for 6 days of simultaneous bimanual two-cursor control as enabled by RNN decoders. Light gray lines connect data points corresponding to the same session day. Each trial had a 10 s timeout after which the trial was considered failed. Angular error was calculated within an initial movement window (300–500 ms after go cue). Vertical black bars in each panel are 95% CIs (bootstrap, n = 10 K). Performance was generally good across most days, although decoders did sometimes fail to enable consistent target acquisition. (b) A sequential unimanual control strategy (moving one cursor at a time; solid black line) was compared to simultaneous bimanual control (dashed gray line) over 2 sessions, of which the median target acquisition times are shown. The sequential unimanual control strategy led to faster target acquisition. (c) RNN decoders were compared to linear decoders on 2 session days. Each point is the median target acquisition time for the corresponding trial type. Solid lines connect points corresponding to the normal bimanual task (consisting of simultaneous dual movements and unimanual single movements). A variation of the task where only unimanual movements were tested (holding the non-active cursor fixed) was used as a control on trial day 1855 to confirm that linear decoders could succeed in a purely unimanual context (dashed lines).