Gaze control for visually guided manipulation

摘要: Human studies have shown that gaze shifts are mostly driven by the task. One explanation is fixations gather information about task relevant properties, where relevance signalled reward. This thesis pursues primarily an engineering science goal to determine what mechanisms a rational decision maker could employ select location optimally, or near given limited and computation time. To do so we formulate characterise three computational models of shifting (implemented on simulated humanoid robot), which use lookahead imagine informational effects possible fixations. Our first model selects most reduces uncertainty in scene (Unc), second maximises expected rewards reducing (Rew+Unc), third gain cumulative reward (Rew+Unc+Gain). We also present integrated account visual search process into Rew+Unc+Gain scheme. secondary concerned with way humans might next location. compare hand-eye coordination timings our previously published human data, provide evidence only incorporate both (Rew+Unc Rew+Unc+Gain) match data.