Abstract
Many real-world tasks such as interception and navigation require model-based prediction to anticipate future conditions for successful planning and execution of motor responses. In these contexts, error is minimized through accurate model acquisition. Learning the correct model requires accruing a rich set of task information. Response errors enhance this knowledge as different strategies are tested with variable success. We hypothesize that error-correcting actions may not only serve to immediately improve performance, as often concluded from sequential trial dependencies in visuomotor behavior, but also enhance visuomotor training via exploration of model space. We tested this hypothesis using a visuomotor task we call the "helicopter task", in which observers pilot a helicopter through a tunnel by controlling its acceleration, avoiding collision with obstacles and the tunnel itself. Successful planning and execution requires model-based prediction of the helicopter’s dynamics and effects of currently applied controls down the tunnel. The dynamics model comprises the rate of acceleration (control exerted), and the rate of falling (no control exerted). As expected, experience improved observers’ piloting ability: they navigated farther through the tunnel without collision. To assess the degree to which they learned and used the dynamics model, two test tasks were interleaved among piloting sessions. In these tasks, observers adjusted the helicopter’s horizontal position at a fixed, but variable vertical distance from a target in order to ensure contact when released with the controls on (upward acceleration) or turned off (downward drift). Performance in these tasks revealed strong initial sequential trial dependency, yielding responses highly-correlated with previous trial errors that reduced overall response bias. With increased training, this sequential dependency gave way to model-based performance revealed by error dependencies on the vertical distance from which the adjusted helicopter was launched. These results support our hypothesis and demonstrate an important role for error-correction in visuomotor model learning.
Meeting abstract presented at VSS 2012